add required files

app.py

@@ -94,7 +94,7 @@ gr_interface = gr.Interface(
     inputs=input_text,
     outputs=[output_audio],
     title="Tango Audio Generator",
-    description="Generate audio using Tango model by providing a text prompt.",
+    description="Generate audio using Tango by providing a text prompt.",
     allow_flagging=False,
         examples=[
         ["A Dog Barking"],

audioldm/__init__.py

@@ -0,0 +1,8 @@
+from .ldm import LatentDiffusion
+from .utils import seed_everything, save_wave, get_time, get_duration
+from .pipeline import *
+
+
+
+
+

audioldm/__main__.py

@@ -0,0 +1,183 @@
+#!/usr/bin/python3
+import os
+from audioldm import text_to_audio, style_transfer, build_model, save_wave, get_time, round_up_duration, get_duration
+import argparse
+
+CACHE_DIR = os.getenv(
+    "AUDIOLDM_CACHE_DIR",
+    os.path.join(os.path.expanduser("~"), ".cache/audioldm"))
+
+parser = argparse.ArgumentParser()
+
+parser.add_argument(
+    "--mode",
+    type=str,
+    required=False,
+    default="generation",
+    help="generation: text-to-audio generation; transfer: style transfer",
+    choices=["generation", "transfer"]
+)
+
+parser.add_argument(
+    "-t",
+    "--text",
+    type=str,
+    required=False,
+    default="",
+    help="Text prompt to the model for audio generation",
+)
+
+parser.add_argument(
+    "-f",
+    "--file_path",
+    type=str,
+    required=False,
+    default=None,
+    help="(--mode transfer): Original audio file for style transfer; Or (--mode generation): the guidance audio file for generating simialr audio",
+)
+
+parser.add_argument(
+    "--transfer_strength",
+    type=float,
+    required=False,
+    default=0.5,
+    help="A value between 0 and 1. 0 means original audio without transfer, 1 means completely transfer to the audio indicated by text",
+)
+
+parser.add_argument(
+    "-s",
+    "--save_path",
+    type=str,
+    required=False,
+    help="The path to save model output",
+    default="./output",
+)
+
+parser.add_argument(
+    "--model_name",
+    type=str,
+    required=False,
+    help="The checkpoint you gonna use",
+    default="audioldm-s-full",
+    choices=["audioldm-s-full", "audioldm-l-full", "audioldm-s-full-v2"]
+)
+
+parser.add_argument(
+    "-ckpt",
+    "--ckpt_path",
+    type=str,
+    required=False,
+    help="The path to the pretrained .ckpt model",
+    default=None,
+)
+
+parser.add_argument(
+    "-b",
+    "--batchsize",
+    type=int,
+    required=False,
+    default=1,
+    help="Generate how many samples at the same time",
+)
+
+parser.add_argument(
+    "--ddim_steps",
+    type=int,
+    required=False,
+    default=200,
+    help="The sampling step for DDIM",
+)
+
+parser.add_argument(
+    "-gs",
+    "--guidance_scale",
+    type=float,
+    required=False,
+    default=2.5,
+    help="Guidance scale (Large => better quality and relavancy to text; Small => better diversity)",
+)
+
+parser.add_argument(
+    "-dur",
+    "--duration",
+    type=float,
+    required=False,
+    default=10.0,
+    help="The duration of the samples",
+)
+
+parser.add_argument(
+    "-n",
+    "--n_candidate_gen_per_text",
+    type=int,
+    required=False,
+    default=3,
+    help="Automatic quality control. This number control the number of candidates (e.g., generate three audios and choose the best to show you). A Larger value usually lead to better quality with heavier computation",
+)
+
+parser.add_argument(
+    "--seed",
+    type=int,
+    required=False,
+    default=42,
+    help="Change this value (any integer number) will lead to a different generation result.",
+)
+
+args = parser.parse_args()
+
+if(args.ckpt_path is not None):
+    print("Warning: ckpt_path has no effect after version 0.0.20.")
+    
+assert args.duration % 2.5 == 0, "Duration must be a multiple of 2.5"
+
+mode = args.mode
+if(mode == "generation" and args.file_path is not None):
+    mode = "generation_audio_to_audio"
+    if(len(args.text) > 0):
+        print("Warning: You have specified the --file_path. --text will be ignored")
+        args.text = ""
+        
+save_path = os.path.join(args.save_path, mode)
+
+if(args.file_path is not None):
+    save_path = os.path.join(save_path, os.path.basename(args.file_path.split(".")[0]))
+
+text = args.text
+random_seed = args.seed
+duration = args.duration
+guidance_scale = args.guidance_scale
+n_candidate_gen_per_text = args.n_candidate_gen_per_text
+
+os.makedirs(save_path, exist_ok=True)
+audioldm = build_model(model_name=args.model_name)
+
+if(args.mode == "generation"):
+    waveform = text_to_audio(
+        audioldm,
+        text,
+        args.file_path,
+        random_seed,
+        duration=duration,
+        guidance_scale=guidance_scale,
+        ddim_steps=args.ddim_steps,
+        n_candidate_gen_per_text=n_candidate_gen_per_text,
+        batchsize=args.batchsize,
+    )
+    
+elif(args.mode == "transfer"):
+    assert args.file_path is not None
+    assert os.path.exists(args.file_path), "The original audio file \'%s\' for style transfer does not exist." % args.file_path
+    waveform = style_transfer(
+        audioldm,
+        text,
+        args.file_path,
+        args.transfer_strength,
+        random_seed,
+        duration=duration,
+        guidance_scale=guidance_scale,
+        ddim_steps=args.ddim_steps,
+        batchsize=args.batchsize,
+    )
+    waveform = waveform[:,None,:]
+
+save_wave(waveform, save_path, name="%s_%s" % (get_time(), text))

audioldm/audio/__init__.py

@@ -0,0 +1,2 @@
+from .tools import wav_to_fbank, read_wav_file
+from .stft import TacotronSTFT

audioldm/audio/audio_processing.py

@@ -0,0 +1,100 @@
+import torch
+import numpy as np
+import librosa.util as librosa_util
+from scipy.signal import get_window
+
+
+def window_sumsquare(
+    window,
+    n_frames,
+    hop_length,
+    win_length,
+    n_fft,
+    dtype=np.float32,
+    norm=None,
+):
+    """
+    # from librosa 0.6
+    Compute the sum-square envelope of a window function at a given hop length.
+
+    This is used to estimate modulation effects induced by windowing
+    observations in short-time fourier transforms.
+
+    Parameters
+    ----------
+    window : string, tuple, number, callable, or list-like
+        Window specification, as in `get_window`
+
+    n_frames : int > 0
+        The number of analysis frames
+
+    hop_length : int > 0
+        The number of samples to advance between frames
+
+    win_length : [optional]
+        The length of the window function.  By default, this matches `n_fft`.
+
+    n_fft : int > 0
+        The length of each analysis frame.
+
+    dtype : np.dtype
+        The data type of the output
+
+    Returns
+    -------
+    wss : np.ndarray, shape=`(n_fft + hop_length * (n_frames - 1))`
+        The sum-squared envelope of the window function
+    """
+    if win_length is None:
+        win_length = n_fft
+
+    n = n_fft + hop_length * (n_frames - 1)
+    x = np.zeros(n, dtype=dtype)
+
+    # Compute the squared window at the desired length
+    win_sq = get_window(window, win_length, fftbins=True)
+    win_sq = librosa_util.normalize(win_sq, norm=norm) ** 2
+    win_sq = librosa_util.pad_center(win_sq, n_fft)
+
+    # Fill the envelope
+    for i in range(n_frames):
+        sample = i * hop_length
+        x[sample : min(n, sample + n_fft)] += win_sq[: max(0, min(n_fft, n - sample))]
+    return x
+
+
+def griffin_lim(magnitudes, stft_fn, n_iters=30):
+    """
+    PARAMS
+    ------
+    magnitudes: spectrogram magnitudes
+    stft_fn: STFT class with transform (STFT) and inverse (ISTFT) methods
+    """
+
+    angles = np.angle(np.exp(2j * np.pi * np.random.rand(*magnitudes.size())))
+    angles = angles.astype(np.float32)
+    angles = torch.autograd.Variable(torch.from_numpy(angles))
+    signal = stft_fn.inverse(magnitudes, angles).squeeze(1)
+
+    for i in range(n_iters):
+        _, angles = stft_fn.transform(signal)
+        signal = stft_fn.inverse(magnitudes, angles).squeeze(1)
+    return signal
+
+
+def dynamic_range_compression(x, normalize_fun=torch.log, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return normalize_fun(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    """
+    PARAMS
+    ------
+    C: compression factor used to compress
+    """
+    return torch.exp(x) / C

audioldm/audio/stft.py

@@ -0,0 +1,186 @@
+import torch
+import torch.nn.functional as F
+import numpy as np
+from scipy.signal import get_window
+from librosa.util import pad_center, tiny
+from librosa.filters import mel as librosa_mel_fn
+
+from audioldm.audio.audio_processing import (
+    dynamic_range_compression,
+    dynamic_range_decompression,
+    window_sumsquare,
+)
+
+
+class STFT(torch.nn.Module):
+    """adapted from Prem Seetharaman's https://github.com/pseeth/pytorch-stft"""
+
+    def __init__(self, filter_length, hop_length, win_length, window="hann"):
+        super(STFT, self).__init__()
+        self.filter_length = filter_length
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.window = window
+        self.forward_transform = None
+        scale = self.filter_length / self.hop_length
+        fourier_basis = np.fft.fft(np.eye(self.filter_length))
+
+        cutoff = int((self.filter_length / 2 + 1))
+        fourier_basis = np.vstack(
+            [np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])]
+        )
+
+        forward_basis = torch.FloatTensor(fourier_basis[:, None, :])
+        inverse_basis = torch.FloatTensor(
+            np.linalg.pinv(scale * fourier_basis).T[:, None, :]
+        )
+
+        if window is not None:
+            assert filter_length >= win_length
+            # get window and zero center pad it to filter_length
+            fft_window = get_window(window, win_length, fftbins=True)
+            fft_window = pad_center(fft_window, filter_length)
+            fft_window = torch.from_numpy(fft_window).float()
+
+            # window the bases
+            forward_basis *= fft_window
+            inverse_basis *= fft_window
+
+        self.register_buffer("forward_basis", forward_basis.float())
+        self.register_buffer("inverse_basis", inverse_basis.float())
+
+    def transform(self, input_data):
+        device = self.forward_basis.device
+        input_data = input_data.to(device)
+        
+        num_batches = input_data.size(0)
+        num_samples = input_data.size(1)
+
+        self.num_samples = num_samples
+
+        # similar to librosa, reflect-pad the input
+        input_data = input_data.view(num_batches, 1, num_samples)
+        input_data = F.pad(
+            input_data.unsqueeze(1),
+            (int(self.filter_length / 2), int(self.filter_length / 2), 0, 0),
+            mode="reflect",
+        )
+        input_data = input_data.squeeze(1)
+
+        forward_transform = F.conv1d(
+            input_data,
+            torch.autograd.Variable(self.forward_basis, requires_grad=False),
+            stride=self.hop_length,
+            padding=0,
+        )#.cpu()
+
+        cutoff = int((self.filter_length / 2) + 1)
+        real_part = forward_transform[:, :cutoff, :]
+        imag_part = forward_transform[:, cutoff:, :]
+
+        magnitude = torch.sqrt(real_part**2 + imag_part**2)
+        phase = torch.autograd.Variable(torch.atan2(imag_part.data, real_part.data))
+
+        return magnitude, phase
+
+    def inverse(self, magnitude, phase):
+        device = self.forward_basis.device
+        magnitude, phase = magnitude.to(device), phase.to(device)
+        
+        recombine_magnitude_phase = torch.cat(
+            [magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
+        )
+
+        inverse_transform = F.conv_transpose1d(
+            recombine_magnitude_phase,
+            torch.autograd.Variable(self.inverse_basis, requires_grad=False),
+            stride=self.hop_length,
+            padding=0,
+        )
+
+        if self.window is not None:
+            window_sum = window_sumsquare(
+                self.window,
+                magnitude.size(-1),
+                hop_length=self.hop_length,
+                win_length=self.win_length,
+                n_fft=self.filter_length,
+                dtype=np.float32,
+            )
+            # remove modulation effects
+            approx_nonzero_indices = torch.from_numpy(
+                np.where(window_sum > tiny(window_sum))[0]
+            )
+            window_sum = torch.autograd.Variable(
+                torch.from_numpy(window_sum), requires_grad=False
+            )
+            window_sum = window_sum
+            inverse_transform[:, :, approx_nonzero_indices] /= window_sum[
+                approx_nonzero_indices
+            ]
+
+            # scale by hop ratio
+            inverse_transform *= float(self.filter_length) / self.hop_length
+
+        inverse_transform = inverse_transform[:, :, int(self.filter_length / 2) :]
+        inverse_transform = inverse_transform[:, :, : -int(self.filter_length / 2) :]
+
+        return inverse_transform
+
+    def forward(self, input_data):
+        self.magnitude, self.phase = self.transform(input_data)
+        reconstruction = self.inverse(self.magnitude, self.phase)
+        return reconstruction
+
+
+class TacotronSTFT(torch.nn.Module):
+    def __init__(
+        self,
+        filter_length,
+        hop_length,
+        win_length,
+        n_mel_channels,
+        sampling_rate,
+        mel_fmin,
+        mel_fmax,
+    ):
+        super(TacotronSTFT, self).__init__()
+        self.n_mel_channels = n_mel_channels
+        self.sampling_rate = sampling_rate
+        self.stft_fn = STFT(filter_length, hop_length, win_length)
+        mel_basis = librosa_mel_fn(
+            sampling_rate, filter_length, n_mel_channels, mel_fmin, mel_fmax
+        )
+        mel_basis = torch.from_numpy(mel_basis).float()
+        self.register_buffer("mel_basis", mel_basis)
+
+    def spectral_normalize(self, magnitudes, normalize_fun):
+        output = dynamic_range_compression(magnitudes, normalize_fun)
+        return output
+
+    def spectral_de_normalize(self, magnitudes):
+        output = dynamic_range_decompression(magnitudes)
+        return output
+
+    def mel_spectrogram(self, y, normalize_fun=torch.log):
+        """Computes mel-spectrograms from a batch of waves
+        PARAMS
+        ------
+        y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
+
+        RETURNS
+        -------
+        mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
+        """
+        assert torch.min(y.data) >= -1, torch.min(y.data)
+        assert torch.max(y.data) <= 1, torch.max(y.data)
+
+        magnitudes, phases = self.stft_fn.transform(y)
+        magnitudes = magnitudes.data
+        mel_output = torch.matmul(self.mel_basis, magnitudes)
+        mel_output = self.spectral_normalize(mel_output, normalize_fun)
+        energy = torch.norm(magnitudes, dim=1)
+
+        log_magnitudes = self.spectral_normalize(magnitudes, normalize_fun)
+
+        return mel_output, log_magnitudes, energy

audioldm/audio/tools.py

@@ -0,0 +1,85 @@
+import torch
+import numpy as np
+import torchaudio
+
+
+def get_mel_from_wav(audio, _stft):
+    audio = torch.clip(torch.FloatTensor(audio).unsqueeze(0), -1, 1)
+    audio = torch.autograd.Variable(audio, requires_grad=False)
+    melspec, log_magnitudes_stft, energy = _stft.mel_spectrogram(audio)
+    melspec = torch.squeeze(melspec, 0).numpy().astype(np.float32)
+    log_magnitudes_stft = (
+        torch.squeeze(log_magnitudes_stft, 0).numpy().astype(np.float32)
+    )
+    energy = torch.squeeze(energy, 0).numpy().astype(np.float32)
+    return melspec, log_magnitudes_stft, energy
+
+
+def _pad_spec(fbank, target_length=1024):
+    n_frames = fbank.shape[0]
+    p = target_length - n_frames
+    # cut and pad
+    if p > 0:
+        m = torch.nn.ZeroPad2d((0, 0, 0, p))
+        fbank = m(fbank)
+    elif p < 0:
+        fbank = fbank[0:target_length, :]
+
+    if fbank.size(-1) % 2 != 0:
+        fbank = fbank[..., :-1]
+
+    return fbank
+
+
+def pad_wav(waveform, segment_length):
+    waveform_length = waveform.shape[-1]
+    assert waveform_length > 100, "Waveform is too short, %s" % waveform_length
+    if segment_length is None or waveform_length == segment_length:
+        return waveform
+    elif waveform_length > segment_length:
+        return waveform[:segment_length]
+    elif waveform_length < segment_length:
+        temp_wav = np.zeros((1, segment_length))
+        temp_wav[:, :waveform_length] = waveform
+    return temp_wav
+
+def normalize_wav(waveform):
+    waveform = waveform - np.mean(waveform)
+    waveform = waveform / (np.max(np.abs(waveform)) + 1e-8)
+    return waveform * 0.5
+
+
+def read_wav_file(filename, segment_length):
+    # waveform, sr = librosa.load(filename, sr=None, mono=True) # 4 times slower
+    waveform, sr = torchaudio.load(filename)  # Faster!!!
+    waveform = torchaudio.functional.resample(waveform, orig_freq=sr, new_freq=16000)
+    waveform = waveform.numpy()[0, ...]
+    waveform = normalize_wav(waveform)
+    waveform = waveform[None, ...]
+    waveform = pad_wav(waveform, segment_length)
+    
+    waveform = waveform / np.max(np.abs(waveform))
+    waveform = 0.5 * waveform
+    
+    return waveform
+
+
+def wav_to_fbank(filename, target_length=1024, fn_STFT=None):
+    assert fn_STFT is not None
+
+    # mixup
+    waveform = read_wav_file(filename, target_length * 160)  # hop size is 160
+
+    waveform = waveform[0, ...]
+    waveform = torch.FloatTensor(waveform)
+
+    fbank, log_magnitudes_stft, energy = get_mel_from_wav(waveform, fn_STFT)
+
+    fbank = torch.FloatTensor(fbank.T)
+    log_magnitudes_stft = torch.FloatTensor(log_magnitudes_stft.T)
+
+    fbank, log_magnitudes_stft = _pad_spec(fbank, target_length), _pad_spec(
+        log_magnitudes_stft, target_length
+    )
+
+    return fbank, log_magnitudes_stft, waveform

audioldm/hifigan/__init__.py

@@ -0,0 +1,7 @@
+from .models import Generator
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self

audioldm/hifigan/models.py

@@ -0,0 +1,174 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils import weight_norm, remove_weight_norm
+
+LRELU_SLOPE = 0.1
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+class ResBlock(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock, self).__init__()
+        self.h = h
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(init_weights)
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class Generator(torch.nn.Module):
+    def __init__(self, h):
+        super(Generator, self).__init__()
+        self.h = h
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+        self.conv_pre = weight_norm(
+            Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)
+        )
+        resblock = ResBlock
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        h.upsample_initial_channel // (2**i),
+                        h.upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(h, ch, k, d))
+
+        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+
+    def forward(self, x):
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        # print("Removing weight norm...")
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)

audioldm/hifigan/utilities.py

@@ -0,0 +1,86 @@
+import os
+import json
+
+import torch
+import numpy as np
+
+import audioldm.hifigan as hifigan
+
+HIFIGAN_16K_64 = {
+    "resblock": "1",
+    "num_gpus": 6,
+    "batch_size": 16,
+    "learning_rate": 0.0002,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.999,
+    "seed": 1234,
+    "upsample_rates": [5, 4, 2, 2, 2],
+    "upsample_kernel_sizes": [16, 16, 8, 4, 4],
+    "upsample_initial_channel": 1024,
+    "resblock_kernel_sizes": [3, 7, 11],
+    "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+    "segment_size": 8192,
+    "num_mels": 64,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 160,
+    "win_size": 1024,
+    "sampling_rate": 16000,
+    "fmin": 0,
+    "fmax": 8000,
+    "fmax_for_loss": None,
+    "num_workers": 4,
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1,
+    },
+}
+
+
+def get_available_checkpoint_keys(model, ckpt):
+    print("==> Attemp to reload from %s" % ckpt)
+    state_dict = torch.load(ckpt)["state_dict"]
+    current_state_dict = model.state_dict()
+    new_state_dict = {}
+    for k in state_dict.keys():
+        if (
+            k in current_state_dict.keys()
+            and current_state_dict[k].size() == state_dict[k].size()
+        ):
+            new_state_dict[k] = state_dict[k]
+        else:
+            print("==> WARNING: Skipping %s" % k)
+    print(
+        "%s out of %s keys are matched"
+        % (len(new_state_dict.keys()), len(state_dict.keys()))
+    )
+    return new_state_dict
+
+
+def get_param_num(model):
+    num_param = sum(param.numel() for param in model.parameters())
+    return num_param
+
+
+def get_vocoder(config, device):
+    config = hifigan.AttrDict(HIFIGAN_16K_64)
+    vocoder = hifigan.Generator(config)
+    vocoder.eval()
+    vocoder.remove_weight_norm()
+    vocoder.to(device)
+    return vocoder
+
+
+def vocoder_infer(mels, vocoder, lengths=None):
+    vocoder.eval()
+    with torch.no_grad():
+        wavs = vocoder(mels).squeeze(1)
+
+    wavs = (wavs.cpu().numpy() * 32768).astype("int16")
+
+    if lengths is not None:
+        wavs = wavs[:, :lengths]
+
+    return wavs

audioldm/latent_diffusion/attention.py

@@ -0,0 +1,469 @@
+from inspect import isfunction
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn
+from einops import rearrange
+
+from audioldm.latent_diffusion.util import checkpoint
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return {el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = (
+            nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
+            if not glu
+            else GEGLU(dim, inner_dim)
+        )
+
+        self.net = nn.Sequential(
+            project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(
+        num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
+    )
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(
+            qkv, "b (qkv heads c) h w -> qkv b heads c (h w)", heads=self.heads, qkv=3
+        )
+        k = k.softmax(dim=-1)
+        context = torch.einsum("bhdn,bhen->bhde", k, v)
+        out = torch.einsum("bhde,bhdn->bhen", context, q)
+        out = rearrange(
+            out, "b heads c (h w) -> b (heads c) h w", heads=self.heads, h=h, w=w
+        )
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = rearrange(q, "b c h w -> b (h w) c")
+        k = rearrange(k, "b c h w -> b c (h w)")
+        w_ = torch.einsum("bij,bjk->bik", q, k)
+
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, "b c h w -> b c (h w)")
+        w_ = rearrange(w_, "b i j -> b j i")
+        h_ = torch.einsum("bij,bjk->bik", v, w_)
+        h_ = rearrange(h_, "b c (h w) -> b c h w", h=h)
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+class CrossAttention(nn.Module):
+    """
+    ### Cross Attention Layer
+    This falls-back to self-attention when conditional embeddings are not specified.
+    """
+
+    # use_flash_attention: bool = True
+    use_flash_attention: bool = False
+
+    def __init__(
+        self,
+        query_dim,
+        context_dim=None,
+        heads=8,
+        dim_head=64,
+        dropout=0.0,
+        is_inplace: bool = True,
+    ):
+        # def __init__(self, d_model: int, d_cond: int, n_heads: int, d_head: int, is_inplace: bool = True):
+        """
+        :param d_model: is the input embedding size
+        :param n_heads: is the number of attention heads
+        :param d_head: is the size of a attention head
+        :param d_cond: is the size of the conditional embeddings
+        :param is_inplace: specifies whether to perform the attention softmax computation inplace to
+            save memory
+        """
+        super().__init__()
+
+        self.is_inplace = is_inplace
+        self.n_heads = heads
+        self.d_head = dim_head
+
+        # Attention scaling factor
+        self.scale = dim_head**-0.5
+
+        # The normal self-attention layer
+        if context_dim is None:
+            context_dim = query_dim
+
+        # Query, key and value mappings
+        d_attn = dim_head * heads
+        self.to_q = nn.Linear(query_dim, d_attn, bias=False)
+        self.to_k = nn.Linear(context_dim, d_attn, bias=False)
+        self.to_v = nn.Linear(context_dim, d_attn, bias=False)
+
+        # Final linear layer
+        self.to_out = nn.Sequential(nn.Linear(d_attn, query_dim), nn.Dropout(dropout))
+
+        # Setup [flash attention](https://github.com/HazyResearch/flash-attention).
+        # Flash attention is only used if it's installed
+        # and `CrossAttention.use_flash_attention` is set to `True`.
+        try:
+            # You can install flash attention by cloning their Github repo,
+            # [https://github.com/HazyResearch/flash-attention](https://github.com/HazyResearch/flash-attention)
+            # and then running `python setup.py install`
+            from flash_attn.flash_attention import FlashAttention
+
+            self.flash = FlashAttention()
+            # Set the scale for scaled dot-product attention.
+            self.flash.softmax_scale = self.scale
+        # Set to `None` if it's not installed
+        except ImportError:
+            self.flash = None
+
+    def forward(self, x, context=None, mask=None):
+        """
+        :param x: are the input embeddings of shape `[batch_size, height * width, d_model]`
+        :param cond: is the conditional embeddings of shape `[batch_size, n_cond, d_cond]`
+        """
+
+        # If `cond` is `None` we perform self attention
+        has_cond = context is not None
+        if not has_cond:
+            context = x
+
+        # Get query, key and value vectors
+        q = self.to_q(x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        # Use flash attention if it's available and the head size is less than or equal to `128`
+        if (
+            CrossAttention.use_flash_attention
+            and self.flash is not None
+            and not has_cond
+            and self.d_head <= 128
+        ):
+            return self.flash_attention(q, k, v)
+        # Otherwise, fallback to normal attention
+        else:
+            return self.normal_attention(q, k, v)
+
+    def flash_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
+        """
+        #### Flash Attention
+        :param q: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+        :param k: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+        :param v: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+        """
+
+        # Get batch size and number of elements along sequence axis (`width * height`)
+        batch_size, seq_len, _ = q.shape
+
+        # Stack `q`, `k`, `v` vectors for flash attention, to get a single tensor of
+        # shape `[batch_size, seq_len, 3, n_heads * d_head]`
+        qkv = torch.stack((q, k, v), dim=2)
+        # Split the heads
+        qkv = qkv.view(batch_size, seq_len, 3, self.n_heads, self.d_head)
+
+        # Flash attention works for head sizes `32`, `64` and `128`, so we have to pad the heads to
+        # fit this size.
+        if self.d_head <= 32:
+            pad = 32 - self.d_head
+        elif self.d_head <= 64:
+            pad = 64 - self.d_head
+        elif self.d_head <= 128:
+            pad = 128 - self.d_head
+        else:
+            raise ValueError(f"Head size ${self.d_head} too large for Flash Attention")
+
+        # Pad the heads
+        if pad:
+            qkv = torch.cat(
+                (qkv, qkv.new_zeros(batch_size, seq_len, 3, self.n_heads, pad)), dim=-1
+            )
+
+        # Compute attention
+        # $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)V$$
+        # This gives a tensor of shape `[batch_size, seq_len, n_heads, d_padded]`
+        # TODO here I add the dtype changing
+        out, _ = self.flash(qkv.type(torch.float16))
+        # Truncate the extra head size
+        out = out[:, :, :, : self.d_head].float()
+        # Reshape to `[batch_size, seq_len, n_heads * d_head]`
+        out = out.reshape(batch_size, seq_len, self.n_heads * self.d_head)
+
+        # Map to `[batch_size, height * width, d_model]` with a linear layer
+        return self.to_out(out)
+
+    def normal_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
+        """
+        #### Normal Attention
+
+        :param q: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+        :param k: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+        :param v: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
+        """
+
+        # Split them to heads of shape `[batch_size, seq_len, n_heads, d_head]`
+        q = q.view(*q.shape[:2], self.n_heads, -1)  # [bs, 64, 20, 32]
+        k = k.view(*k.shape[:2], self.n_heads, -1)  # [bs, 1, 20, 32]
+        v = v.view(*v.shape[:2], self.n_heads, -1)
+
+        # Calculate attention $\frac{Q K^\top}{\sqrt{d_{key}}}$
+        attn = torch.einsum("bihd,bjhd->bhij", q, k) * self.scale
+
+        # Compute softmax
+        # $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)$$
+        if self.is_inplace:
+            half = attn.shape[0] // 2
+            attn[half:] = attn[half:].softmax(dim=-1)
+            attn[:half] = attn[:half].softmax(dim=-1)
+        else:
+            attn = attn.softmax(dim=-1)
+
+        # Compute attention output
+        # $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)V$$
+        # attn: [bs, 20, 64, 1]
+        # v: [bs, 1, 20, 32]
+        out = torch.einsum("bhij,bjhd->bihd", attn, v)
+        # Reshape to `[batch_size, height * width, n_heads * d_head]`
+        out = out.reshape(*out.shape[:2], -1)
+        # Map to `[batch_size, height * width, d_model]` with a linear layer
+        return self.to_out(out)
+
+
+# class CrossAttention(nn.Module):
+# def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
+#     super().__init__()
+#     inner_dim = dim_head * heads
+#     context_dim = default(context_dim, query_dim)
+
+#     self.scale = dim_head ** -0.5
+#     self.heads = heads
+
+#     self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+#     self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+#     self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+#     self.to_out = nn.Sequential(
+#         nn.Linear(inner_dim, query_dim),
+#         nn.Dropout(dropout)
+#     )
+
+# def forward(self, x, context=None, mask=None):
+#     h = self.heads
+
+#     q = self.to_q(x)
+#     context = default(context, x)
+#     k = self.to_k(context)
+#     v = self.to_v(context)
+
+#     q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+#     sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+
+#     if exists(mask):
+#         mask = rearrange(mask, 'b ... -> b (...)')
+#         max_neg_value = -torch.finfo(sim.dtype).max
+#         mask = repeat(mask, 'b j -> (b h) () j', h=h)
+#         sim.masked_fill_(~mask, max_neg_value)
+
+#     # attention, what we cannot get enough of
+#     attn = sim.softmax(dim=-1)
+
+#     out = einsum('b i j, b j d -> b i d', attn, v)
+#     out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+#     return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        dropout=0.0,
+        context_dim=None,
+        gated_ff=True,
+        checkpoint=True,
+    ):
+        super().__init__()
+        self.attn1 = CrossAttention(
+            query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout
+        )  # is a self-attention
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = CrossAttention(
+            query_dim=dim,
+            context_dim=context_dim,
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+        )  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        if context is None:
+            return checkpoint(self._forward, (x,), self.parameters(), self.checkpoint)
+        else:
+            return checkpoint(
+                self._forward, (x, context), self.parameters(), self.checkpoint
+            )
+
+    def _forward(self, x, context=None):
+        x = self.attn1(self.norm1(x)) + x
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        n_heads,
+        d_head,
+        depth=1,
+        dropout=0.0,
+        context_dim=None,
+        no_context=False,
+    ):
+        super().__init__()
+
+        if no_context:
+            context_dim = None
+
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+
+        self.proj_in = nn.Conv2d(
+            in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim
+                )
+                for d in range(depth)
+            ]
+        )
+
+        self.proj_out = zero_module(
+            nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+        )
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = self.proj_in(x)
+        x = rearrange(x, "b c h w -> b (h w) c")
+        for block in self.transformer_blocks:
+            x = block(x, context=context)
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
+        x = self.proj_out(x)
+        return x + x_in

audioldm/latent_diffusion/ddim.py

@@ -0,0 +1,377 @@
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+
+from audioldm.latent_diffusion.util import (
+    make_ddim_sampling_parameters,
+    make_ddim_timesteps,
+    noise_like,
+    extract_into_tensor,
+)
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(
+        self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0.0, verbose=True
+    ):
+        self.ddim_timesteps = make_ddim_timesteps(
+            ddim_discr_method=ddim_discretize,
+            num_ddim_timesteps=ddim_num_steps,
+            num_ddpm_timesteps=self.ddpm_num_timesteps,
+            verbose=verbose,
+        )
+        alphas_cumprod = self.model.alphas_cumprod
+        assert (
+            alphas_cumprod.shape[0] == self.ddpm_num_timesteps
+        ), "alphas have to be defined for each timestep"
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer("betas", to_torch(self.model.betas))
+        self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod))
+        self.register_buffer(
+            "alphas_cumprod_prev", to_torch(self.model.alphas_cumprod_prev)
+        )
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer(
+            "sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod.cpu()))
+        )
+        self.register_buffer(
+            "sqrt_one_minus_alphas_cumprod",
+            to_torch(np.sqrt(1.0 - alphas_cumprod.cpu())),
+        )
+        self.register_buffer(
+            "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod.cpu()))
+        )
+        self.register_buffer(
+            "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod.cpu()))
+        )
+        self.register_buffer(
+            "sqrt_recipm1_alphas_cumprod",
+            to_torch(np.sqrt(1.0 / alphas_cumprod.cpu() - 1)),
+        )
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(
+            alphacums=alphas_cumprod.cpu(),
+            ddim_timesteps=self.ddim_timesteps,
+            eta=ddim_eta,
+            verbose=verbose,
+        )
+        self.register_buffer("ddim_sigmas", ddim_sigmas)
+        self.register_buffer("ddim_alphas", ddim_alphas)
+        self.register_buffer("ddim_alphas_prev", ddim_alphas_prev)
+        self.register_buffer("ddim_sqrt_one_minus_alphas", np.sqrt(1.0 - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev)
+            / (1 - self.alphas_cumprod)
+            * (1 - self.alphas_cumprod / self.alphas_cumprod_prev)
+        )
+        self.register_buffer(
+            "ddim_sigmas_for_original_num_steps", sigmas_for_original_sampling_steps
+        )
+
+    @torch.no_grad()
+    def sample(
+        self,
+        S,
+        batch_size,
+        shape,
+        conditioning=None,
+        callback=None,
+        normals_sequence=None,
+        img_callback=None,
+        quantize_x0=False,
+        eta=0.0,
+        mask=None,
+        x0=None,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        verbose=True,
+        x_T=None,
+        log_every_t=100,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+        **kwargs,
+    ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                if cbs != batch_size:
+                    print(
+                        f"Warning: Got {cbs} conditionings but batch-size is {batch_size}"
+                    )
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(
+                        f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}"
+                    )
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        samples, intermediates = self.ddim_sampling(
+            conditioning,
+            size,
+            callback=callback,
+            img_callback=img_callback,
+            quantize_denoised=quantize_x0,
+            mask=mask,
+            x0=x0,
+            ddim_use_original_steps=False,
+            noise_dropout=noise_dropout,
+            temperature=temperature,
+            score_corrector=score_corrector,
+            corrector_kwargs=corrector_kwargs,
+            x_T=x_T,
+            log_every_t=log_every_t,
+            unconditional_guidance_scale=unconditional_guidance_scale,
+            unconditional_conditioning=unconditional_conditioning,
+        )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(
+        self,
+        cond,
+        shape,
+        x_T=None,
+        ddim_use_original_steps=False,
+        callback=None,
+        timesteps=None,
+        quantize_denoised=False,
+        mask=None,
+        x0=None,
+        img_callback=None,
+        log_every_t=100,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+    ):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = (
+                self.ddpm_num_timesteps
+                if ddim_use_original_steps
+                else self.ddim_timesteps
+            )
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = (
+                int(
+                    min(timesteps / self.ddim_timesteps.shape[0], 1)
+                    * self.ddim_timesteps.shape[0]
+                )
+                - 1
+            )
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {"x_inter": [img], "pred_x0": [img]}
+        time_range = (
+            reversed(range(0, timesteps))
+            if ddim_use_original_steps
+            else np.flip(timesteps)
+        )
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        # print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        # iterator = gr.Progress().tqdm(time_range, desc="DDIM Sampler", total=total_steps)
+        iterator = tqdm(time_range, desc="DDIM Sampler", total=total_steps, leave=False)
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(
+                    x0, ts
+                )  # TODO deterministic forward pass?
+                img = (
+                    img_orig * mask + (1.0 - mask) * img
+                )  # In the first sampling step, img is pure gaussian noise
+
+            outs = self.p_sample_ddim(
+                img,
+                cond,
+                ts,
+                index=index,
+                use_original_steps=ddim_use_original_steps,
+                quantize_denoised=quantize_denoised,
+                temperature=temperature,
+                noise_dropout=noise_dropout,
+                score_corrector=score_corrector,
+                corrector_kwargs=corrector_kwargs,
+                unconditional_guidance_scale=unconditional_guidance_scale,
+                unconditional_conditioning=unconditional_conditioning,
+            )
+            img, pred_x0 = outs
+            if callback:
+                callback(i)
+            if img_callback:
+                img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates["x_inter"].append(img)
+                intermediates["pred_x0"].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+
+        return (
+            extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0
+            + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise
+        )
+
+    @torch.no_grad()
+    def decode(
+        self,
+        x_latent,
+        cond,
+        t_start,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        use_original_steps=False,
+    ):
+
+        timesteps = (
+            np.arange(self.ddpm_num_timesteps)
+            if use_original_steps
+            else self.ddim_timesteps
+        )
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        # print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        # iterator = gr.Progress().tqdm(time_range, desc="Decoding image", total=total_steps)
+        iterator = tqdm(time_range, desc="Decoding image", total=total_steps)
+        x_dec = x_latent
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full(
+                (x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long
+            )
+            x_dec, _ = self.p_sample_ddim(
+                x_dec,
+                cond,
+                ts,
+                index=index,
+                use_original_steps=use_original_steps,
+                unconditional_guidance_scale=unconditional_guidance_scale,
+                unconditional_conditioning=unconditional_conditioning,
+            )
+        return x_dec
+
+    @torch.no_grad()
+    def p_sample_ddim(
+        self,
+        x,
+        c,
+        t,
+        index,
+        repeat_noise=False,
+        use_original_steps=False,
+        quantize_denoised=False,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+    ):
+        b, *_, device = *x.shape, x.device
+
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.0:
+            e_t = self.model.apply_model(x, t, c)
+        else:
+            x_in = torch.cat([x] * 2)
+            t_in = torch.cat([t] * 2)
+            c_in = torch.cat([unconditional_conditioning, c])
+            e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+            # When unconditional_guidance_scale == 1: only e_t
+            # When unconditional_guidance_scale == 0: only unconditional
+            # When unconditional_guidance_scale > 1: add more unconditional guidance
+            e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps"
+            e_t = score_corrector.modify_score(
+                self.model, e_t, x, t, c, **corrector_kwargs
+            )
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = (
+            self.model.alphas_cumprod_prev
+            if use_original_steps
+            else self.ddim_alphas_prev
+        )
+        sqrt_one_minus_alphas = (
+            self.model.sqrt_one_minus_alphas_cumprod
+            if use_original_steps
+            else self.ddim_sqrt_one_minus_alphas
+        )
+        sigmas = (
+            self.model.ddim_sigmas_for_original_num_steps
+            if use_original_steps
+            else self.ddim_sigmas
+        )
+        # select parameters corresponding to the currently considered timestep
+        a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+        a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+        sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full(
+            (b, 1, 1, 1), sqrt_one_minus_alphas[index], device=device
+        )
+
+        # current prediction for x_0
+        pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+        # direction pointing to x_t
+        dir_xt = (1.0 - a_prev - sigma_t**2).sqrt() * e_t
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.0:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise  # TODO
+        return x_prev, pred_x0

audioldm/latent_diffusion/ddpm.py

@@ -0,0 +1,441 @@
+"""
+wild mixture of
+https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
+https://github.com/CompVis/taming-transformers
+-- merci
+"""
+import sys
+import os
+
+import torch
+import torch.nn as nn
+import numpy as np
+from contextlib import contextmanager
+from functools import partial
+from tqdm import tqdm
+
+from audioldm.utils import exists, default, count_params, instantiate_from_config
+from audioldm.latent_diffusion.ema import LitEma
+from audioldm.latent_diffusion.util import (
+    make_beta_schedule,
+    extract_into_tensor,
+    noise_like,
+)
+import soundfile as sf
+import os
+
+
+__conditioning_keys__ = {"concat": "c_concat", "crossattn": "c_crossattn", "adm": "y"}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+def uniform_on_device(r1, r2, shape, device):
+    return (r1 - r2) * torch.rand(*shape, device=device) + r2
+
+
+class DiffusionWrapper(nn.Module):
+    def __init__(self, diff_model_config, conditioning_key):
+        super().__init__()
+        self.diffusion_model = instantiate_from_config(diff_model_config)
+        self.conditioning_key = conditioning_key
+        assert self.conditioning_key in [
+            None,
+            "concat",
+            "crossattn",
+            "hybrid",
+            "adm",
+            "film",
+        ]
+
+    def forward(
+        self, x, t, c_concat: list = None, c_crossattn: list = None, c_film: list = None
+    ):
+        x = x.contiguous()
+        t = t.contiguous()
+
+        if self.conditioning_key is None:
+            out = self.diffusion_model(x, t)
+        elif self.conditioning_key == "concat":
+            xc = torch.cat([x] + c_concat, dim=1)
+            out = self.diffusion_model(xc, t)
+        elif self.conditioning_key == "crossattn":
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(x, t, context=cc)
+        elif self.conditioning_key == "hybrid":
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc)
+        elif (
+            self.conditioning_key == "film"
+        ):  # The condition is assumed to be a global token, which wil pass through a linear layer and added with the time embedding for the FILM
+            cc = c_film[0].squeeze(1)  # only has one token
+            out = self.diffusion_model(x, t, y=cc)
+        elif self.conditioning_key == "adm":
+            cc = c_crossattn[0]
+            out = self.diffusion_model(x, t, y=cc)
+        else:
+            raise NotImplementedError()
+
+        return out
+
+
+class DDPM(nn.Module):
+    # classic DDPM with Gaussian diffusion, in image space
+    def __init__(
+        self,
+        unet_config,
+        timesteps=1000,
+        beta_schedule="linear",
+        loss_type="l2",
+        ckpt_path=None,
+        ignore_keys=[],
+        load_only_unet=False,
+        monitor="val/loss",
+        use_ema=True,
+        first_stage_key="image",
+        latent_t_size=256,
+        latent_f_size=16,
+        channels=3,
+        log_every_t=100,
+        clip_denoised=True,
+        linear_start=1e-4,
+        linear_end=2e-2,
+        cosine_s=8e-3,
+        given_betas=None,
+        original_elbo_weight=0.0,
+        v_posterior=0.0,  # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
+        l_simple_weight=1.0,
+        conditioning_key=None,
+        parameterization="eps",  # all assuming fixed variance schedules
+        scheduler_config=None,
+        use_positional_encodings=False,
+        learn_logvar=False,
+        logvar_init=0.0,
+    ):
+        super().__init__()
+        assert parameterization in [
+            "eps",
+            "x0",
+        ], 'currently only supporting "eps" and "x0"'
+        self.parameterization = parameterization
+        self.state = None
+        # print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
+        self.cond_stage_model = None
+        self.clip_denoised = clip_denoised
+        self.log_every_t = log_every_t
+        self.first_stage_key = first_stage_key
+
+        self.latent_t_size = latent_t_size
+        self.latent_f_size = latent_f_size
+
+        self.channels = channels
+        self.use_positional_encodings = use_positional_encodings
+        self.model = DiffusionWrapper(unet_config, conditioning_key)
+        count_params(self.model, verbose=True)
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self.model)
+            # print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        self.use_scheduler = scheduler_config is not None
+        if self.use_scheduler:
+            self.scheduler_config = scheduler_config
+
+        self.v_posterior = v_posterior
+        self.original_elbo_weight = original_elbo_weight
+        self.l_simple_weight = l_simple_weight
+
+        if monitor is not None:
+            self.monitor = monitor
+
+        self.register_schedule(
+            given_betas=given_betas,
+            beta_schedule=beta_schedule,
+            timesteps=timesteps,
+            linear_start=linear_start,
+            linear_end=linear_end,
+            cosine_s=cosine_s,
+        )
+
+        self.loss_type = loss_type
+
+        self.learn_logvar = learn_logvar
+        self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
+        if self.learn_logvar:
+            self.logvar = nn.Parameter(self.logvar, requires_grad=True)
+        else:
+            self.logvar = nn.Parameter(self.logvar, requires_grad=False)
+
+        self.logger_save_dir = None
+        self.logger_project = None
+        self.logger_version = None
+        self.label_indices_total = None
+        # To avoid the system cannot find metric value for checkpoint
+        self.metrics_buffer = {
+            "val/kullback_leibler_divergence_sigmoid": 15.0,
+            "val/kullback_leibler_divergence_softmax": 10.0,
+            "val/psnr": 0.0,
+            "val/ssim": 0.0,
+            "val/inception_score_mean": 1.0,
+            "val/inception_score_std": 0.0,
+            "val/kernel_inception_distance_mean": 0.0,
+            "val/kernel_inception_distance_std": 0.0,
+            "val/frechet_inception_distance": 133.0,
+            "val/frechet_audio_distance": 32.0,
+        }
+        self.initial_learning_rate = None
+
+    def get_log_dir(self):
+        if (
+            self.logger_save_dir is None
+            and self.logger_project is None
+            and self.logger_version is None
+        ):
+            return os.path.join(
+                self.logger.save_dir, self.logger._project, self.logger.version
+            )
+        else:
+            return os.path.join(
+                self.logger_save_dir, self.logger_project, self.logger_version
+            )
+
+    def set_log_dir(self, save_dir, project, version):
+        self.logger_save_dir = save_dir
+        self.logger_project = project
+        self.logger_version = version
+
+    def register_schedule(
+        self,
+        given_betas=None,
+        beta_schedule="linear",
+        timesteps=1000,
+        linear_start=1e-4,
+        linear_end=2e-2,
+        cosine_s=8e-3,
+    ):
+        if exists(given_betas):
+            betas = given_betas
+        else:
+            betas = make_beta_schedule(
+                beta_schedule,
+                timesteps,
+                linear_start=linear_start,
+                linear_end=linear_end,
+                cosine_s=cosine_s,
+            )
+        alphas = 1.0 - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1.0, alphas_cumprod[:-1])
+
+        (timesteps,) = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert (
+            alphas_cumprod.shape[0] == self.num_timesteps
+        ), "alphas have to be defined for each timestep"
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer("betas", to_torch(betas))
+        self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod))
+        self.register_buffer("alphas_cumprod_prev", to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer("sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer(
+            "sqrt_one_minus_alphas_cumprod", to_torch(np.sqrt(1.0 - alphas_cumprod))
+        )
+        self.register_buffer(
+            "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod))
+        )
+        self.register_buffer(
+            "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod))
+        )
+        self.register_buffer(
+            "sqrt_recipm1_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod - 1))
+        )
+
+        # calculations for posterior q(x_{t-1} | x_t, x_0)
+        posterior_variance = (1 - self.v_posterior) * betas * (
+            1.0 - alphas_cumprod_prev
+        ) / (1.0 - alphas_cumprod) + self.v_posterior * betas
+        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
+        self.register_buffer("posterior_variance", to_torch(posterior_variance))
+        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
+        self.register_buffer(
+            "posterior_log_variance_clipped",
+            to_torch(np.log(np.maximum(posterior_variance, 1e-20))),
+        )
+        self.register_buffer(
+            "posterior_mean_coef1",
+            to_torch(betas * np.sqrt(alphas_cumprod_prev) / (1.0 - alphas_cumprod)),
+        )
+        self.register_buffer(
+            "posterior_mean_coef2",
+            to_torch(
+                (1.0 - alphas_cumprod_prev) * np.sqrt(alphas) / (1.0 - alphas_cumprod)
+            ),
+        )
+
+        if self.parameterization == "eps":
+            lvlb_weights = self.betas**2 / (
+                2
+                * self.posterior_variance
+                * to_torch(alphas)
+                * (1 - self.alphas_cumprod)
+            )
+        elif self.parameterization == "x0":
+            lvlb_weights = (
+                0.5
+                * np.sqrt(torch.Tensor(alphas_cumprod))
+                / (2.0 * 1 - torch.Tensor(alphas_cumprod))
+            )
+        else:
+            raise NotImplementedError("mu not supported")
+        # TODO how to choose this term
+        lvlb_weights[0] = lvlb_weights[1]
+        self.register_buffer("lvlb_weights", lvlb_weights, persistent=False)
+        assert not torch.isnan(self.lvlb_weights).all()
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.model.parameters())
+            self.model_ema.copy_to(self.model)
+            if context is not None:
+                # print(f"{context}: Switched to EMA weights")
+                pass
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.model.parameters())
+                if context is not None:
+                    # print(f"{context}: Restored training weights")
+                    pass
+
+    def q_mean_variance(self, x_start, t):
+        """
+        Get the distribution q(x_t | x_0).
+        :param x_start: the [N x C x ...] tensor of noiseless inputs.
+        :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
+        :return: A tuple (mean, variance, log_variance), all of x_start's shape.
+        """
+        mean = extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
+        variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
+        log_variance = extract_into_tensor(
+            self.log_one_minus_alphas_cumprod, t, x_start.shape
+        )
+        return mean, variance, log_variance
+
+    def predict_start_from_noise(self, x_t, t, noise):
+        return (
+            extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t
+            - extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+            * noise
+        )
+
+    def q_posterior(self, x_start, x_t, t):
+        posterior_mean = (
+            extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start
+            + extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
+        )
+        posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
+        posterior_log_variance_clipped = extract_into_tensor(
+            self.posterior_log_variance_clipped, t, x_t.shape
+        )
+        return posterior_mean, posterior_variance, posterior_log_variance_clipped
+
+    def p_mean_variance(self, x, t, clip_denoised: bool):
+        model_out = self.model(x, t)
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        if clip_denoised:
+            x_recon.clamp_(-1.0, 1.0)
+
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(
+            x_start=x_recon, x_t=x, t=t
+        )
+        return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
+        b, *_, device = *x.shape, x.device
+        model_mean, _, model_log_variance = self.p_mean_variance(
+            x=x, t=t, clip_denoised=clip_denoised
+        )
+        noise = noise_like(x.shape, device, repeat_noise)
+        # no noise when t == 0
+        nonzero_mask = (
+            (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))).contiguous()
+        )
+        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, shape, return_intermediates=False):
+        device = self.betas.device
+        b = shape[0]
+        img = torch.randn(shape, device=device)
+        intermediates = [img]
+        for i in tqdm(
+            reversed(range(0, self.num_timesteps)),
+            desc="Sampling t",
+            total=self.num_timesteps,
+        ):
+            img = self.p_sample(
+                img,
+                torch.full((b,), i, device=device, dtype=torch.long),
+                clip_denoised=self.clip_denoised,
+            )
+            if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
+                intermediates.append(img)
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, batch_size=16, return_intermediates=False):
+        shape = (batch_size, channels, self.latent_t_size, self.latent_f_size)
+        channels = self.channels
+        return self.p_sample_loop(shape, return_intermediates=return_intermediates)
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (
+            extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
+            + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape)
+            * noise
+        )
+
+    def forward(self, x, *args, **kwargs):
+        t = torch.randint(
+            0, self.num_timesteps, (x.shape[0],), device=self.device
+        ).long()
+        return self.p_losses(x, t, *args, **kwargs)
+
+    def get_input(self, batch, k):
+        # fbank, log_magnitudes_stft, label_indices, fname, waveform, clip_label, text = batch
+        fbank, log_magnitudes_stft, label_indices, fname, waveform, text = batch
+        ret = {}
+
+        ret["fbank"] = (
+            fbank.unsqueeze(1).to(memory_format=torch.contiguous_format).float()
+        )
+        ret["stft"] = log_magnitudes_stft.to(
+            memory_format=torch.contiguous_format
+        ).float()
+        # ret["clip_label"] = clip_label.to(memory_format=torch.contiguous_format).float()
+        ret["waveform"] = waveform.to(memory_format=torch.contiguous_format).float()
+        ret["text"] = list(text)
+        ret["fname"] = fname
+
+        return ret[k]

audioldm/latent_diffusion/ema.py

@@ -0,0 +1,82 @@
+import torch
+from torch import nn
+
+
+class LitEma(nn.Module):
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError("Decay must be between 0 and 1")
+
+        self.m_name2s_name = {}
+        self.register_buffer("decay", torch.tensor(decay, dtype=torch.float32))
+        self.register_buffer(
+            "num_updates",
+            torch.tensor(0, dtype=torch.int)
+            if use_num_upates
+            else torch.tensor(-1, dtype=torch.int),
+        )
+
+        for name, p in model.named_parameters():
+            if p.requires_grad:
+                # remove as '.'-character is not allowed in buffers
+                s_name = name.replace(".", "")
+                self.m_name2s_name.update({name: s_name})
+                self.register_buffer(s_name, p.clone().detach().data)
+
+        self.collected_params = []
+
+    def forward(self, model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with torch.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if m_param[key].requires_grad:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
+                    shadow_params[sname].sub_(
+                        one_minus_decay * (shadow_params[sname] - m_param[key])
+                    )
+                else:
+                    assert not key in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if m_param[key].requires_grad:
+                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
+            else:
+                assert not key in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            param.data.copy_(c_param.data)

audioldm/latent_diffusion/openaimodel.py

@@ -0,0 +1,1069 @@
+from abc import abstractmethod
+import math
+
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+
+from audioldm.latent_diffusion.util import (
+    checkpoint,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    zero_module,
+    normalization,
+    timestep_embedding,
+)
+from audioldm.latent_diffusion.attention import SpatialTransformer
+
+
+# dummy replace
+def convert_module_to_f16(x):
+    pass
+
+
+def convert_module_to_f32(x):
+    pass
+
+
+## go
+class AttentionPool2d(nn.Module):
+    """
+    Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
+    """
+
+    def __init__(
+        self,
+        spacial_dim: int,
+        embed_dim: int,
+        num_heads_channels: int,
+        output_dim: int = None,
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(
+            th.randn(embed_dim, spacial_dim**2 + 1) / embed_dim**0.5
+        )
+        self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
+        self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
+        self.num_heads = embed_dim // num_heads_channels
+        self.attention = QKVAttention(self.num_heads)
+
+    def forward(self, x):
+        b, c, *_spatial = x.shape
+        x = x.reshape(b, c, -1).contiguous()  # NC(HW)
+        x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1)  # NC(HW+1)
+        x = x + self.positional_embedding[None, :, :].to(x.dtype)  # NC(HW+1)
+        x = self.qkv_proj(x)
+        x = self.attention(x)
+        x = self.c_proj(x)
+        return x[:, :, 0]
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(
+                dims, self.channels, self.out_channels, 3, padding=padding
+            )
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(
+                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
+            )
+        else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class TransposedUpsample(nn.Module):
+    "Learned 2x upsampling without padding"
+
+    def __init__(self, channels, out_channels=None, ks=5):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+
+        self.up = nn.ConvTranspose2d(
+            self.channels, self.out_channels, kernel_size=ks, stride=2
+        )
+
+    def forward(self, x):
+        return self.up(x)
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims,
+                self.channels,
+                self.out_channels,
+                3,
+                stride=stride,
+                padding=padding,
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, 3, padding=1
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        return checkpoint(
+            self._forward, (x, emb), self.parameters(), self.use_checkpoint
+        )
+
+    def _forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = th.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    An attention block that allows spatial positions to attend to each other.
+    Originally ported from here, but adapted to the N-d case.
+    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
+    """
+
+    def __init__(
+        self,
+        channels,
+        num_heads=1,
+        num_head_channels=-1,
+        use_checkpoint=False,
+        use_new_attention_order=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        if num_head_channels == -1:
+            self.num_heads = num_heads
+        else:
+            assert (
+                channels % num_head_channels == 0
+            ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
+            self.num_heads = channels // num_head_channels
+        self.use_checkpoint = use_checkpoint
+        self.norm = normalization(channels)
+        self.qkv = conv_nd(1, channels, channels * 3, 1)
+        if use_new_attention_order:
+            # split qkv before split heads
+            self.attention = QKVAttention(self.num_heads)
+        else:
+            # split heads before split qkv
+            self.attention = QKVAttentionLegacy(self.num_heads)
+
+        self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
+
+    def forward(self, x):
+        return checkpoint(
+            self._forward, (x,), self.parameters(), True
+        )  # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
+        # return pt_checkpoint(self._forward, x)  # pytorch
+
+    def _forward(self, x):
+        b, c, *spatial = x.shape
+        x = x.reshape(b, c, -1).contiguous()
+        qkv = self.qkv(self.norm(x)).contiguous()
+        h = self.attention(qkv).contiguous()
+        h = self.proj_out(h).contiguous()
+        return (x + h).reshape(b, c, *spatial).contiguous()
+
+
+def count_flops_attn(model, _x, y):
+    """
+    A counter for the `thop` package to count the operations in an
+    attention operation.
+    Meant to be used like:
+        macs, params = thop.profile(
+            model,
+            inputs=(inputs, timestamps),
+            custom_ops={QKVAttention: QKVAttention.count_flops},
+        )
+    """
+    b, c, *spatial = y[0].shape
+    num_spatial = int(np.prod(spatial))
+    # We perform two matmuls with the same number of ops.
+    # The first computes the weight matrix, the second computes
+    # the combination of the value vectors.
+    matmul_ops = 2 * b * (num_spatial**2) * c
+    model.total_ops += th.DoubleTensor([matmul_ops])
+
+
+class QKVAttentionLegacy(nn.Module):
+    """
+    A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = (
+            qkv.reshape(bs * self.n_heads, ch * 3, length).contiguous().split(ch, dim=1)
+        )
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts", q * scale, k * scale
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v)
+        return a.reshape(bs, -1, length).contiguous()
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class QKVAttention(nn.Module):
+    """
+    A module which performs QKV attention and splits in a different order.
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.chunk(3, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts",
+            (q * scale).view(bs * self.n_heads, ch, length),
+            (k * scale).view(bs * self.n_heads, ch, length),
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum(
+            "bts,bcs->bct",
+            weight,
+            v.reshape(bs * self.n_heads, ch, length).contiguous(),
+        )
+        return a.reshape(bs, -1, length).contiguous()
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        extra_film_condition_dim=None,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        extra_film_use_concat=False,  # If true, concatenate extrafilm condition with time embedding, else addition
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,  # custom transformer support
+        transformer_depth=1,  # custom transformer support
+        context_dim=None,  # custom transformer support
+        n_embed=None,  # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+    ):
+        super().__init__()
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert (
+                num_head_channels != -1
+            ), "Either num_heads or num_head_channels has to be set"
+
+        if num_head_channels == -1:
+            assert (
+                num_heads != -1
+            ), "Either num_heads or num_head_channels has to be set"
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.extra_film_condition_dim = extra_film_condition_dim
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+        self.extra_film_use_concat = extra_film_use_concat
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        assert not (
+            self.num_classes is not None and self.extra_film_condition_dim is not None
+        ), "As for the condition of theh UNet model, you can only set using class label or an extra embedding vector (such as from CLAP). You cannot set both num_classes and extra_film_condition_dim."
+
+        if self.num_classes is not None:
+            self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+
+        self.use_extra_film_by_concat = (
+            self.extra_film_condition_dim is not None and self.extra_film_use_concat
+        )
+        self.use_extra_film_by_addition = (
+            self.extra_film_condition_dim is not None and not self.extra_film_use_concat
+        )
+
+        if self.extra_film_condition_dim is not None:
+            self.film_emb = nn.Linear(self.extra_film_condition_dim, time_embed_dim)
+            # print("+ Use extra condition on UNet channel using Film. Extra condition dimension is %s. " % self.extra_film_condition_dim)
+            # if(self.use_extra_film_by_concat):
+            #     print("\t By concatenation with time embedding")
+            # elif(self.use_extra_film_by_concat):
+            #     print("\t By addition with time embedding")
+
+        if use_spatial_transformer and (
+            self.use_extra_film_by_concat or self.use_extra_film_by_addition
+        ):
+            # print("+ Spatial transformer will only be used as self-attention. Because you have choose to use film as your global condition.")
+            spatial_transformer_no_context = True
+        else:
+            spatial_transformer_no_context = False
+
+        if use_spatial_transformer and not spatial_transformer_no_context:
+            assert (
+                context_dim is not None
+            ), "Fool!! You forgot to include the dimension of your cross-attention conditioning..."
+
+        if context_dim is not None and not spatial_transformer_no_context:
+            assert (
+                use_spatial_transformer
+            ), "Fool!! You forgot to use the spatial transformer for your cross-attention conditioning..."
+            from omegaconf.listconfig import ListConfig
+
+            if type(context_dim) == ListConfig:
+                context_dim = list(context_dim)
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim
+                        if (not self.use_extra_film_by_concat)
+                        else time_embed_dim * 2,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        dim_head = (
+                            ch // num_heads
+                            if use_spatial_transformer
+                            else num_head_channels
+                        )
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=dim_head,
+                            use_new_attention_order=use_new_attention_order,
+                        )
+                        if not use_spatial_transformer
+                        else SpatialTransformer(
+                            ch,
+                            num_heads,
+                            dim_head,
+                            depth=transformer_depth,
+                            context_dim=context_dim,
+                            no_context=spatial_transformer_no_context,
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim
+                            if (not self.use_extra_film_by_concat)
+                            else time_embed_dim * 2,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            # num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim
+                if (not self.use_extra_film_by_concat)
+                else time_embed_dim * 2,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                use_new_attention_order=use_new_attention_order,
+            )
+            if not use_spatial_transformer
+            else SpatialTransformer(
+                ch,
+                num_heads,
+                dim_head,
+                depth=transformer_depth,
+                context_dim=context_dim,
+                no_context=spatial_transformer_no_context,
+            ),
+            ResBlock(
+                ch,
+                time_embed_dim
+                if (not self.use_extra_film_by_concat)
+                else time_embed_dim * 2,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(num_res_blocks + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim
+                        if (not self.use_extra_film_by_concat)
+                        else time_embed_dim * 2,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        # num_heads = 1
+                        dim_head = (
+                            ch // num_heads
+                            if use_spatial_transformer
+                            else num_head_channels
+                        )
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads_upsample,
+                            num_head_channels=dim_head,
+                            use_new_attention_order=use_new_attention_order,
+                        )
+                        if not use_spatial_transformer
+                        else SpatialTransformer(
+                            ch,
+                            num_heads,
+                            dim_head,
+                            depth=transformer_depth,
+                            context_dim=context_dim,
+                            no_context=spatial_transformer_no_context,
+                        )
+                    )
+                if level and i == num_res_blocks:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim
+                            if (not self.use_extra_film_by_concat)
+                            else time_embed_dim * 2,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+                normalization(ch),
+                conv_nd(dims, model_channels, n_embed, 1),
+                # nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+            )
+
+        self.shape_reported = False
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+        self.output_blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+        self.output_blocks.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps=None, context=None, y=None, **kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional. an [N, extra_film_condition_dim] Tensor if film-embed conditional
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        if not self.shape_reported:
+            # print("The shape of UNet input is", x.size())
+            self.shape_reported = True
+
+        assert (y is not None) == (
+            self.num_classes is not None or self.extra_film_condition_dim is not None
+        ), "must specify y if and only if the model is class-conditional or film embedding conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+
+        if self.num_classes is not None:
+            assert y.shape == (x.shape[0],)
+            emb = emb + self.label_emb(y)
+
+        if self.use_extra_film_by_addition:
+            emb = emb + self.film_emb(y)
+        elif self.use_extra_film_by_concat:
+            emb = th.cat([emb, self.film_emb(y)], dim=-1)
+
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb, context)
+            hs.append(h)
+        h = self.middle_block(h, emb, context)
+        for module in self.output_blocks:
+            h = th.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)
+
+
+class EncoderUNetModel(nn.Module):
+    """
+    The half UNet model with attention and timestep embedding.
+    For usage, see UNet.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        pool="adaptive",
+        *args,
+        **kwargs,
+    ):
+        super().__init__()
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=num_head_channels,
+                            use_new_attention_order=use_new_attention_order,
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=num_head_channels,
+                use_new_attention_order=use_new_attention_order,
+            ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+        self.pool = pool
+        if pool == "adaptive":
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                nn.AdaptiveAvgPool2d((1, 1)),
+                zero_module(conv_nd(dims, ch, out_channels, 1)),
+                nn.Flatten(),
+            )
+        elif pool == "attention":
+            assert num_head_channels != -1
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                AttentionPool2d(
+                    (image_size // ds), ch, num_head_channels, out_channels
+                ),
+            )
+        elif pool == "spatial":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                nn.ReLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        elif pool == "spatial_v2":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                normalization(2048),
+                nn.SiLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        else:
+            raise NotImplementedError(f"Unexpected {pool} pooling")
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :return: an [N x K] Tensor of outputs.
+        """
+        emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
+        results = []
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb)
+            if self.pool.startswith("spatial"):
+                results.append(h.type(x.dtype).mean(dim=(2, 3)))
+        h = self.middle_block(h, emb)
+        if self.pool.startswith("spatial"):
+            results.append(h.type(x.dtype).mean(dim=(2, 3)))
+            h = th.cat(results, axis=-1)
+            return self.out(h)
+        else:
+            h = h.type(x.dtype)
+            return self.out(h)

audioldm/latent_diffusion/util.py

@@ -0,0 +1,295 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+
+import os
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import repeat
+
+from audioldm.utils import instantiate_from_config
+
+
+def make_beta_schedule(
+    schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3
+):
+    if schedule == "linear":
+        betas = (
+            torch.linspace(
+                linear_start**0.5, linear_end**0.5, n_timestep, dtype=torch.float64
+            )
+            ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+            torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(
+            linear_start, linear_end, n_timestep, dtype=torch.float64
+        )
+    elif schedule == "sqrt":
+        betas = (
+            torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+            ** 0.5
+        )
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+
+def make_ddim_timesteps(
+    ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True
+):
+    if ddim_discr_method == "uniform":
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+    elif ddim_discr_method == "quad":
+        ddim_timesteps = (
+            (np.linspace(0, np.sqrt(num_ddpm_timesteps * 0.8), num_ddim_timesteps)) ** 2
+        ).astype(int)
+    else:
+        raise NotImplementedError(
+            f'There is no ddim discretization method called "{ddim_discr_method}"'
+        )
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    steps_out = ddim_timesteps + 1
+    if verbose:
+        print(f"Selected timesteps for ddim sampler: {steps_out}")
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt(
+        (1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)
+    )
+    if verbose:
+        print(
+            f"Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}"
+        )
+        print(
+            f"For the chosen value of eta, which is {eta}, "
+            f"this results in the following sigma_t schedule for ddim sampler {sigmas}"
+        )
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t).contiguous()
+    return out.reshape(b, *((1,) * (len(x_shape) - 1))).contiguous()
+
+
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad():
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period)
+            * torch.arange(start=0, end=half, dtype=torch.float32)
+            / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat(
+                [embedding, torch.zeros_like(embedding[:, :1])], dim=-1
+            )
+    else:
+        embedding = repeat(timesteps, "b -> b d", d=dim)
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def normalization(channels):
+    """
+    Make a standard normalization layer.
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    return GroupNorm32(32, channels)
+
+
+# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
+class SiLU(nn.Module):
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class GroupNorm32(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.float()).type(x.dtype)
+
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+class HybridConditioner(nn.Module):
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {"c_concat": [c_concat], "c_crossattn": [c_crossattn]}
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(
+        shape[0], *((1,) * (len(shape) - 1))
+    )
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()

audioldm/ldm.py

@@ -0,0 +1,818 @@
+import os
+
+import torch
+import numpy as np
+from tqdm import tqdm
+from audioldm.utils import default, instantiate_from_config, save_wave
+from audioldm.latent_diffusion.ddpm import DDPM
+from audioldm.variational_autoencoder.distributions import DiagonalGaussianDistribution
+from audioldm.latent_diffusion.util import noise_like
+from audioldm.latent_diffusion.ddim import DDIMSampler
+import os
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class LatentDiffusion(DDPM):
+    """main class"""
+
+    def __init__(
+        self,
+        device="cuda",
+        first_stage_config=None,
+        cond_stage_config=None,
+        num_timesteps_cond=None,
+        cond_stage_key="image",
+        cond_stage_trainable=False,
+        concat_mode=True,
+        cond_stage_forward=None,
+        conditioning_key=None,
+        scale_factor=1.0,
+        scale_by_std=False,
+        base_learning_rate=None,
+        *args,
+        **kwargs,
+    ):
+        self.device = device
+        self.learning_rate = base_learning_rate
+        self.num_timesteps_cond = default(num_timesteps_cond, 1)
+        self.scale_by_std = scale_by_std
+        assert self.num_timesteps_cond <= kwargs["timesteps"]
+        # for backwards compatibility after implementation of DiffusionWrapper
+        if conditioning_key is None:
+            conditioning_key = "concat" if concat_mode else "crossattn"
+        if cond_stage_config == "__is_unconditional__":
+            conditioning_key = None
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", [])
+        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
+        self.concat_mode = concat_mode
+        self.cond_stage_trainable = cond_stage_trainable
+        self.cond_stage_key = cond_stage_key
+        self.cond_stage_key_orig = cond_stage_key
+        try:
+            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
+        except:
+            self.num_downs = 0
+        if not scale_by_std:
+            self.scale_factor = scale_factor
+        else:
+            self.register_buffer("scale_factor", torch.tensor(scale_factor))
+        self.instantiate_first_stage(first_stage_config)
+        self.instantiate_cond_stage(cond_stage_config)
+        self.cond_stage_forward = cond_stage_forward
+        self.clip_denoised = False
+
+    def make_cond_schedule(
+        self,
+    ):
+        self.cond_ids = torch.full(
+            size=(self.num_timesteps,),
+            fill_value=self.num_timesteps - 1,
+            dtype=torch.long,
+        )
+        ids = torch.round(
+            torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)
+        ).long()
+        self.cond_ids[: self.num_timesteps_cond] = ids
+
+    def register_schedule(
+        self,
+        given_betas=None,
+        beta_schedule="linear",
+        timesteps=1000,
+        linear_start=1e-4,
+        linear_end=2e-2,
+        cosine_s=8e-3,
+    ):
+        super().register_schedule(
+            given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s
+        )
+
+        self.shorten_cond_schedule = self.num_timesteps_cond > 1
+        if self.shorten_cond_schedule:
+            self.make_cond_schedule()
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+    def instantiate_cond_stage(self, config):
+        if not self.cond_stage_trainable:
+            if config == "__is_first_stage__":
+                print("Using first stage also as cond stage.")
+                self.cond_stage_model = self.first_stage_model
+            elif config == "__is_unconditional__":
+                print(f"Training {self.__class__.__name__} as an unconditional model.")
+                self.cond_stage_model = None
+                # self.be_unconditional = True
+            else:
+                model = instantiate_from_config(config)
+                self.cond_stage_model = model.eval()
+                self.cond_stage_model.train = disabled_train
+                for param in self.cond_stage_model.parameters():
+                    param.requires_grad = False
+        else:
+            assert config != "__is_first_stage__"
+            assert config != "__is_unconditional__"
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model
+        self.cond_stage_model = self.cond_stage_model.to(self.device)
+
+    def get_first_stage_encoding(self, encoder_posterior):
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample()
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(
+                f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented"
+            )
+        return self.scale_factor * z
+
+    def get_learned_conditioning(self, c):
+        if self.cond_stage_forward is None:
+            if hasattr(self.cond_stage_model, "encode") and callable(
+                self.cond_stage_model.encode
+            ):
+                c = self.cond_stage_model.encode(c)
+                if isinstance(c, DiagonalGaussianDistribution):
+                    c = c.mode()
+            else:
+                # Text input is list
+                if type(c) == list and len(c) == 1:
+                    c = self.cond_stage_model([c[0], c[0]])
+                    c = c[0:1]
+                else:
+                    c = self.cond_stage_model(c)
+        else:
+            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+        return c
+
+    @torch.no_grad()
+    def get_input(
+        self,
+        batch,
+        k,
+        return_first_stage_encode=True,
+        return_first_stage_outputs=False,
+        force_c_encode=False,
+        cond_key=None,
+        return_original_cond=False,
+        bs=None,
+    ):
+        x = super().get_input(batch, k)
+
+        if bs is not None:
+            x = x[:bs]
+
+        x = x.to(self.device)
+
+        if return_first_stage_encode:
+            encoder_posterior = self.encode_first_stage(x)
+            z = self.get_first_stage_encoding(encoder_posterior).detach()
+        else:
+            z = None
+
+        if self.model.conditioning_key is not None:
+            if cond_key is None:
+                cond_key = self.cond_stage_key
+            if cond_key != self.first_stage_key:
+                if cond_key in ["caption", "coordinates_bbox"]:
+                    xc = batch[cond_key]
+                elif cond_key == "class_label":
+                    xc = batch
+                else:
+                    # [bs, 1, 527]
+                    xc = super().get_input(batch, cond_key)
+                    if type(xc) == torch.Tensor:
+                        xc = xc.to(self.device)
+            else:
+                xc = x
+            if not self.cond_stage_trainable or force_c_encode:
+                if isinstance(xc, dict) or isinstance(xc, list):
+                    c = self.get_learned_conditioning(xc)
+                else:
+                    c = self.get_learned_conditioning(xc.to(self.device))
+            else:
+                c = xc
+
+            if bs is not None:
+                c = c[:bs]
+
+        else:
+            c = None
+            xc = None
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                c = {"pos_x": pos_x, "pos_y": pos_y}
+        out = [z, c]
+        if return_first_stage_outputs:
+            xrec = self.decode_first_stage(z)
+            out.extend([x, xrec])
+        if return_original_cond:
+            out.append(xc)
+        return out
+
+    @torch.no_grad()
+    def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, "b h w c -> b c h w").contiguous()
+
+        z = 1.0 / self.scale_factor * z
+        return self.first_stage_model.decode(z)
+
+    def mel_spectrogram_to_waveform(self, mel):
+        # Mel: [bs, 1, t-steps, fbins]
+        if len(mel.size()) == 4:
+            mel = mel.squeeze(1)
+        mel = mel.permute(0, 2, 1)
+        waveform = self.first_stage_model.vocoder(mel)
+        waveform = waveform.cpu().detach().numpy()
+        return waveform
+
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        return self.first_stage_model.encode(x)
+
+    def apply_model(self, x_noisy, t, cond, return_ids=False):
+
+        if isinstance(cond, dict):
+            # hybrid case, cond is exptected to be a dict
+            pass
+        else:
+            if not isinstance(cond, list):
+                cond = [cond]
+            if self.model.conditioning_key == "concat":
+                key = "c_concat"
+            elif self.model.conditioning_key == "crossattn":
+                key = "c_crossattn"
+            else:
+                key = "c_film"
+
+            cond = {key: cond}
+
+        x_recon = self.model(x_noisy, t, **cond)
+
+        if isinstance(x_recon, tuple) and not return_ids:
+            return x_recon[0]
+        else:
+            return x_recon
+
+    def p_mean_variance(
+        self,
+        x,
+        c,
+        t,
+        clip_denoised: bool,
+        return_codebook_ids=False,
+        quantize_denoised=False,
+        return_x0=False,
+        score_corrector=None,
+        corrector_kwargs=None,
+    ):
+        t_in = t
+        model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
+
+        if score_corrector is not None:
+            assert self.parameterization == "eps"
+            model_out = score_corrector.modify_score(
+                self, model_out, x, t, c, **corrector_kwargs
+            )
+
+        if return_codebook_ids:
+            model_out, logits = model_out
+
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        else:
+            raise NotImplementedError()
+
+        if clip_denoised:
+            x_recon.clamp_(-1.0, 1.0)
+        if quantize_denoised:
+            x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(
+            x_start=x_recon, x_t=x, t=t
+        )
+        if return_codebook_ids:
+            return model_mean, posterior_variance, posterior_log_variance, logits
+        elif return_x0:
+            return model_mean, posterior_variance, posterior_log_variance, x_recon
+        else:
+            return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(
+        self,
+        x,
+        c,
+        t,
+        clip_denoised=False,
+        repeat_noise=False,
+        return_codebook_ids=False,
+        quantize_denoised=False,
+        return_x0=False,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+    ):
+        b, *_, device = *x.shape, x.device
+        outputs = self.p_mean_variance(
+            x=x,
+            c=c,
+            t=t,
+            clip_denoised=clip_denoised,
+            return_codebook_ids=return_codebook_ids,
+            quantize_denoised=quantize_denoised,
+            return_x0=return_x0,
+            score_corrector=score_corrector,
+            corrector_kwargs=corrector_kwargs,
+        )
+        if return_codebook_ids:
+            raise DeprecationWarning("Support dropped.")
+            model_mean, _, model_log_variance, logits = outputs
+        elif return_x0:
+            model_mean, _, model_log_variance, x0 = outputs
+        else:
+            model_mean, _, model_log_variance = outputs
+
+        noise = noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.0:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        # no noise when t == 0
+        nonzero_mask = (
+            (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))).contiguous()
+        )
+
+        if return_codebook_ids:
+            return model_mean + nonzero_mask * (
+                0.5 * model_log_variance
+            ).exp() * noise, logits.argmax(dim=1)
+        if return_x0:
+            return (
+                model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise,
+                x0,
+            )
+        else:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def progressive_denoising(
+        self,
+        cond,
+        shape,
+        verbose=True,
+        callback=None,
+        quantize_denoised=False,
+        img_callback=None,
+        mask=None,
+        x0=None,
+        temperature=1.0,
+        noise_dropout=0.0,
+        score_corrector=None,
+        corrector_kwargs=None,
+        batch_size=None,
+        x_T=None,
+        start_T=None,
+        log_every_t=None,
+    ):
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        timesteps = self.num_timesteps
+        if batch_size is not None:
+            b = batch_size if batch_size is not None else shape[0]
+            shape = [batch_size] + list(shape)
+        else:
+            b = batch_size = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=self.device)
+        else:
+            img = x_T
+        intermediates = []
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {
+                    key: cond[key][:batch_size]
+                    if not isinstance(cond[key], list)
+                    else list(map(lambda x: x[:batch_size], cond[key]))
+                    for key in cond
+                }
+            else:
+                cond = (
+                    [c[:batch_size] for c in cond]
+                    if isinstance(cond, list)
+                    else cond[:batch_size]
+                )
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = (
+            tqdm(
+                reversed(range(0, timesteps)),
+                desc="Progressive Generation",
+                total=timesteps,
+            )
+            if verbose
+            else reversed(range(0, timesteps))
+        )
+        if type(temperature) == float:
+            temperature = [temperature] * timesteps
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=self.device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != "hybrid"
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img, x0_partial = self.p_sample(
+                img,
+                cond,
+                ts,
+                clip_denoised=self.clip_denoised,
+                quantize_denoised=quantize_denoised,
+                return_x0=True,
+                temperature=temperature[i],
+                noise_dropout=noise_dropout,
+                score_corrector=score_corrector,
+                corrector_kwargs=corrector_kwargs,
+            )
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1.0 - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(x0_partial)
+            if callback:
+                callback(i)
+            if img_callback:
+                img_callback(img, i)
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_loop(
+        self,
+        cond,
+        shape,
+        return_intermediates=False,
+        x_T=None,
+        verbose=True,
+        callback=None,
+        timesteps=None,
+        quantize_denoised=False,
+        mask=None,
+        x0=None,
+        img_callback=None,
+        start_T=None,
+        log_every_t=None,
+    ):
+
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        device = self.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        intermediates = [img]
+        if timesteps is None:
+            timesteps = self.num_timesteps
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = (
+            tqdm(reversed(range(0, timesteps)), desc="Sampling t", total=timesteps)
+            if verbose
+            else reversed(range(0, timesteps))
+        )
+
+        if mask is not None:
+            assert x0 is not None
+            assert x0.shape[2:3] == mask.shape[2:3]  # spatial size has to match
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != "hybrid"
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img = self.p_sample(
+                img,
+                cond,
+                ts,
+                clip_denoised=self.clip_denoised,
+                quantize_denoised=quantize_denoised,
+            )
+            if mask is not None:
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1.0 - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(img)
+            if callback:
+                callback(i)
+            if img_callback:
+                img_callback(img, i)
+
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(
+        self,
+        cond,
+        batch_size=16,
+        return_intermediates=False,
+        x_T=None,
+        verbose=True,
+        timesteps=None,
+        quantize_denoised=False,
+        mask=None,
+        x0=None,
+        shape=None,
+        **kwargs,
+    ):
+        if shape is None:
+            shape = (batch_size, self.channels, self.latent_t_size, self.latent_f_size)
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {
+                    key: cond[key][:batch_size]
+                    if not isinstance(cond[key], list)
+                    else list(map(lambda x: x[:batch_size], cond[key]))
+                    for key in cond
+                }
+            else:
+                cond = (
+                    [c[:batch_size] for c in cond]
+                    if isinstance(cond, list)
+                    else cond[:batch_size]
+                )
+        return self.p_sample_loop(
+            cond,
+            shape,
+            return_intermediates=return_intermediates,
+            x_T=x_T,
+            verbose=verbose,
+            timesteps=timesteps,
+            quantize_denoised=quantize_denoised,
+            mask=mask,
+            x0=x0,
+            **kwargs,
+        )
+
+    @torch.no_grad()
+    def sample_log(
+        self,
+        cond,
+        batch_size,
+        ddim,
+        ddim_steps,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        use_plms=False,
+        mask=None,
+        **kwargs,
+    ):
+
+        if mask is not None:
+            shape = (self.channels, mask.size()[-2], mask.size()[-1])
+        else:
+            shape = (self.channels, self.latent_t_size, self.latent_f_size)
+
+        intermediate = None
+        if ddim and not use_plms:
+            # print("Use ddim sampler")
+
+            ddim_sampler = DDIMSampler(self)
+            samples, intermediates = ddim_sampler.sample(
+                ddim_steps,
+                batch_size,
+                shape,
+                cond,
+                verbose=False,
+                unconditional_guidance_scale=unconditional_guidance_scale,
+                unconditional_conditioning=unconditional_conditioning,
+                mask=mask,
+                **kwargs,
+            )
+
+        else:
+            # print("Use DDPM sampler")
+            samples, intermediates = self.sample(
+                cond=cond,
+                batch_size=batch_size,
+                return_intermediates=True,
+                unconditional_guidance_scale=unconditional_guidance_scale,
+                mask=mask,
+                unconditional_conditioning=unconditional_conditioning,
+                **kwargs,
+            )
+
+        return samples, intermediate
+
+    @torch.no_grad()
+    def generate_sample(
+        self,
+        batchs,
+        ddim_steps=200,
+        ddim_eta=1.0,
+        x_T=None,
+        n_candidate_gen_per_text=1,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        name="waveform",
+        use_plms=False,
+        save=False,
+        **kwargs,
+    ):
+        # Generate n_candidate_gen_per_text times and select the best
+        # Batch: audio, text, fnames
+        assert x_T is None
+        try:
+            batchs = iter(batchs)
+        except TypeError:
+            raise ValueError("The first input argument should be an iterable object")
+
+        if use_plms:
+            assert ddim_steps is not None
+        use_ddim = ddim_steps is not None
+        # waveform_save_path = os.path.join(self.get_log_dir(), name)
+        # os.makedirs(waveform_save_path, exist_ok=True)
+        # print("Waveform save path: ", waveform_save_path)
+
+        with self.ema_scope("Generate"):
+            for batch in batchs:
+                z, c = self.get_input(
+                    batch,
+                    self.first_stage_key,
+                    cond_key=self.cond_stage_key,
+                    return_first_stage_outputs=False,
+                    force_c_encode=True,
+                    return_original_cond=False,
+                    bs=None,
+                )
+                text = super().get_input(batch, "text")
+
+                # Generate multiple samples
+                batch_size = z.shape[0] * n_candidate_gen_per_text
+                c = torch.cat([c] * n_candidate_gen_per_text, dim=0)
+                text = text * n_candidate_gen_per_text
+
+                if unconditional_guidance_scale != 1.0:
+                    unconditional_conditioning = (
+                        self.cond_stage_model.get_unconditional_condition(batch_size)
+                    )
+
+                samples, _ = self.sample_log(
+                    cond=c,
+                    batch_size=batch_size,
+                    x_T=x_T,
+                    ddim=use_ddim,
+                    ddim_steps=ddim_steps,
+                    eta=ddim_eta,
+                    unconditional_guidance_scale=unconditional_guidance_scale,
+                    unconditional_conditioning=unconditional_conditioning,
+                    use_plms=use_plms,
+                )
+                
+                if(torch.max(torch.abs(samples)) > 1e2):
+                    samples = torch.clip(samples, min=-10, max=10)
+                    
+                mel = self.decode_first_stage(samples)
+
+                waveform = self.mel_spectrogram_to_waveform(mel)
+
+                if waveform.shape[0] > 1:
+                    similarity = self.cond_stage_model.cos_similarity(
+                        torch.FloatTensor(waveform).squeeze(1), text
+                    )
+
+                    best_index = []
+                    for i in range(z.shape[0]):
+                        candidates = similarity[i :: z.shape[0]]
+                        max_index = torch.argmax(candidates).item()
+                        best_index.append(i + max_index * z.shape[0])
+
+                    waveform = waveform[best_index]
+                    # print("Similarity between generated audio and text", similarity)
+                    # print("Choose the following indexes:", best_index)
+
+        return waveform
+
+    @torch.no_grad()
+    def generate_sample_masked(
+        self,
+        batchs,
+        ddim_steps=200,
+        ddim_eta=1.0,
+        x_T=None,
+        n_candidate_gen_per_text=1,
+        unconditional_guidance_scale=1.0,
+        unconditional_conditioning=None,
+        name="waveform",
+        use_plms=False,
+        time_mask_ratio_start_and_end=(0.25, 0.75),
+        freq_mask_ratio_start_and_end=(0.75, 1.0),
+        save=False,
+        **kwargs,
+    ):
+        # Generate n_candidate_gen_per_text times and select the best
+        # Batch: audio, text, fnames
+        assert x_T is None
+        try:
+            batchs = iter(batchs)
+        except TypeError:
+            raise ValueError("The first input argument should be an iterable object")
+
+        if use_plms:
+            assert ddim_steps is not None
+        use_ddim = ddim_steps is not None
+        # waveform_save_path = os.path.join(self.get_log_dir(), name)
+        # os.makedirs(waveform_save_path, exist_ok=True)
+        # print("Waveform save path: ", waveform_save_path)
+
+        with self.ema_scope("Generate"):
+            for batch in batchs:
+                z, c = self.get_input(
+                    batch,
+                    self.first_stage_key,
+                    cond_key=self.cond_stage_key,
+                    return_first_stage_outputs=False,
+                    force_c_encode=True,
+                    return_original_cond=False,
+                    bs=None,
+                )
+                text = super().get_input(batch, "text")
+                
+                # Generate multiple samples
+                batch_size = z.shape[0] * n_candidate_gen_per_text
+                
+                _, h, w = z.shape[0], z.shape[2], z.shape[3]
+                
+                mask = torch.ones(batch_size, h, w).to(self.device)
+                
+                mask[:, int(h * time_mask_ratio_start_and_end[0]) : int(h * time_mask_ratio_start_and_end[1]), :] = 0 
+                mask[:, :, int(w * freq_mask_ratio_start_and_end[0]) : int(w * freq_mask_ratio_start_and_end[1])] = 0 
+                mask = mask[:, None, ...]
+                
+                c = torch.cat([c] * n_candidate_gen_per_text, dim=0)
+                text = text * n_candidate_gen_per_text
+
+                if unconditional_guidance_scale != 1.0:
+                    unconditional_conditioning = (
+                        self.cond_stage_model.get_unconditional_condition(batch_size)
+                    )
+
+                samples, _ = self.sample_log(
+                    cond=c,
+                    batch_size=batch_size,
+                    x_T=x_T,
+                    ddim=use_ddim,
+                    ddim_steps=ddim_steps,
+                    eta=ddim_eta,
+                    unconditional_guidance_scale=unconditional_guidance_scale,
+                    unconditional_conditioning=unconditional_conditioning,
+                    use_plms=use_plms, mask=mask, x0=torch.cat([z] * n_candidate_gen_per_text)
+                )
+
+                mel = self.decode_first_stage(samples)
+
+                waveform = self.mel_spectrogram_to_waveform(mel)
+
+                if waveform.shape[0] > 1:
+                    similarity = self.cond_stage_model.cos_similarity(
+                        torch.FloatTensor(waveform).squeeze(1), text
+                    )
+
+                    best_index = []
+                    for i in range(z.shape[0]):
+                        candidates = similarity[i :: z.shape[0]]
+                        max_index = torch.argmax(candidates).item()
+                        best_index.append(i + max_index * z.shape[0])
+
+                    waveform = waveform[best_index]
+                    # print("Similarity between generated audio and text", similarity)
+                    # print("Choose the following indexes:", best_index)
+
+        return waveform

audioldm/pipeline.py

@@ -0,0 +1,301 @@
+import os
+
+import argparse
+import yaml
+import torch
+from torch import autocast
+from tqdm import tqdm, trange
+
+from audioldm import LatentDiffusion, seed_everything
+from audioldm.utils import default_audioldm_config, get_duration, get_bit_depth, get_metadata, download_checkpoint
+from audioldm.audio import wav_to_fbank, TacotronSTFT, read_wav_file
+from audioldm.latent_diffusion.ddim import DDIMSampler
+from einops import repeat
+import os
+
+def make_batch_for_text_to_audio(text, waveform=None, fbank=None, batchsize=1):
+    text = [text] * batchsize
+    if batchsize < 1:
+        print("Warning: Batchsize must be at least 1. Batchsize is set to .")
+    
+    if(fbank is None):
+        fbank = torch.zeros((batchsize, 1024, 64))  # Not used, here to keep the code format
+    else:
+        fbank = torch.FloatTensor(fbank)
+        fbank = fbank.expand(batchsize, 1024, 64)
+        assert fbank.size(0) == batchsize
+        
+    stft = torch.zeros((batchsize, 1024, 512))  # Not used
+
+    if(waveform is None):
+        waveform = torch.zeros((batchsize, 160000))  # Not used
+    else:
+        waveform = torch.FloatTensor(waveform)
+        waveform = waveform.expand(batchsize, -1)
+        assert waveform.size(0) == batchsize
+        
+    fname = [""] * batchsize  # Not used
+    
+    batch = (
+        fbank,
+        stft,
+        None,
+        fname,
+        waveform,
+        text,
+    )
+    return batch
+
+def round_up_duration(duration):
+    return int(round(duration/2.5) + 1) * 2.5
+
+def build_model(
+    ckpt_path=None,
+    config=None,
+    model_name="audioldm-s-full"
+):
+    print("Load AudioLDM: %s", model_name)
+    
+    if(ckpt_path is None):
+        ckpt_path = get_metadata()[model_name]["path"]
+    
+    if(not os.path.exists(ckpt_path)):
+        download_checkpoint(model_name)
+
+    if torch.cuda.is_available():
+        device = torch.device("cuda:0")
+    else:
+        device = torch.device("cpu")
+
+    if config is not None:
+        assert type(config) is str
+        config = yaml.load(open(config, "r"), Loader=yaml.FullLoader)
+    else:
+        config = default_audioldm_config(model_name)
+
+    # Use text as condition instead of using waveform during training
+    config["model"]["params"]["device"] = device
+    config["model"]["params"]["cond_stage_key"] = "text"
+
+    # No normalization here
+    latent_diffusion = LatentDiffusion(**config["model"]["params"])
+
+    resume_from_checkpoint = ckpt_path
+
+    checkpoint = torch.load(resume_from_checkpoint, map_location=device)
+    latent_diffusion.load_state_dict(checkpoint["state_dict"])
+
+    latent_diffusion.eval()
+    latent_diffusion = latent_diffusion.to(device)
+
+    latent_diffusion.cond_stage_model.embed_mode = "text"
+    return latent_diffusion
+
+def duration_to_latent_t_size(duration):
+    return int(duration * 25.6)
+
+def set_cond_audio(latent_diffusion):
+    latent_diffusion.cond_stage_key = "waveform"
+    latent_diffusion.cond_stage_model.embed_mode="audio"
+    return latent_diffusion
+
+def set_cond_text(latent_diffusion):
+    latent_diffusion.cond_stage_key = "text"
+    latent_diffusion.cond_stage_model.embed_mode="text"
+    return latent_diffusion
+    
+def text_to_audio(
+    latent_diffusion,
+    text,
+    original_audio_file_path = None,
+    seed=42,
+    ddim_steps=200,
+    duration=10,
+    batchsize=1,
+    guidance_scale=2.5,
+    n_candidate_gen_per_text=3,
+    config=None,
+):
+    seed_everything(int(seed))
+    waveform = None
+    if(original_audio_file_path is not None):
+        waveform = read_wav_file(original_audio_file_path, int(duration * 102.4) * 160)
+        
+    batch = make_batch_for_text_to_audio(text, waveform=waveform, batchsize=batchsize)
+
+    latent_diffusion.latent_t_size = duration_to_latent_t_size(duration)
+    
+    if(waveform is not None):
+        print("Generate audio that has similar content as %s" % original_audio_file_path)
+        latent_diffusion = set_cond_audio(latent_diffusion)
+    else:
+        print("Generate audio using text %s" % text)
+        latent_diffusion = set_cond_text(latent_diffusion)
+        
+    with torch.no_grad():
+        waveform = latent_diffusion.generate_sample(
+            [batch],
+            unconditional_guidance_scale=guidance_scale,
+            ddim_steps=ddim_steps,
+            n_candidate_gen_per_text=n_candidate_gen_per_text,
+            duration=duration,
+        )
+    return waveform
+
+def style_transfer(
+    latent_diffusion,
+    text,
+    original_audio_file_path,
+    transfer_strength,
+    seed=42,
+    duration=10,
+    batchsize=1,
+    guidance_scale=2.5,
+    ddim_steps=200,
+    config=None,
+):
+    if torch.cuda.is_available():
+        device = torch.device("cuda:0")
+    else:
+        device = torch.device("cpu")
+
+    assert original_audio_file_path is not None, "You need to provide the original audio file path"
+    
+    audio_file_duration = get_duration(original_audio_file_path)
+    
+    assert get_bit_depth(original_audio_file_path) == 16, "The bit depth of the original audio file %s must be 16" % original_audio_file_path
+    
+    # if(duration > 20):
+    #     print("Warning: The duration of the audio file %s must be less than 20 seconds. Longer duration will result in Nan in model output (we are still debugging that); Automatically set duration to 20 seconds")
+    #     duration = 20
+    
+    if(duration >= audio_file_duration):
+        print("Warning: Duration you specified %s-seconds must equal or smaller than the audio file duration %ss" % (duration, audio_file_duration))
+        duration = round_up_duration(audio_file_duration)
+        print("Set new duration as %s-seconds" % duration)
+
+    # duration = round_up_duration(duration)
+    
+    latent_diffusion = set_cond_text(latent_diffusion)
+
+    if config is not None:
+        assert type(config) is str
+        config = yaml.load(open(config, "r"), Loader=yaml.FullLoader)
+    else:
+        config = default_audioldm_config()
+
+    seed_everything(int(seed))
+    # latent_diffusion.latent_t_size = duration_to_latent_t_size(duration)
+    latent_diffusion.cond_stage_model.embed_mode = "text"
+
+    fn_STFT = TacotronSTFT(
+        config["preprocessing"]["stft"]["filter_length"],
+        config["preprocessing"]["stft"]["hop_length"],
+        config["preprocessing"]["stft"]["win_length"],
+        config["preprocessing"]["mel"]["n_mel_channels"],
+        config["preprocessing"]["audio"]["sampling_rate"],
+        config["preprocessing"]["mel"]["mel_fmin"],
+        config["preprocessing"]["mel"]["mel_fmax"],
+    )
+
+    mel, _, _ = wav_to_fbank(
+        original_audio_file_path, target_length=int(duration * 102.4), fn_STFT=fn_STFT
+    )
+    mel = mel.unsqueeze(0).unsqueeze(0).to(device)
+    mel = repeat(mel, "1 ... -> b ...", b=batchsize)
+    init_latent = latent_diffusion.get_first_stage_encoding(
+        latent_diffusion.encode_first_stage(mel)
+    )  # move to latent space, encode and sample
+    if(torch.max(torch.abs(init_latent)) > 1e2):
+        init_latent = torch.clip(init_latent, min=-10, max=10)
+    sampler = DDIMSampler(latent_diffusion)
+    sampler.make_schedule(ddim_num_steps=ddim_steps, ddim_eta=1.0, verbose=False)
+
+    t_enc = int(transfer_strength * ddim_steps)
+    prompts = text
+
+    with torch.no_grad():
+        with autocast("cuda"):
+            with latent_diffusion.ema_scope():
+                uc = None
+                if guidance_scale != 1.0:
+                    uc = latent_diffusion.cond_stage_model.get_unconditional_condition(
+                        batchsize
+                    )
+
+                c = latent_diffusion.get_learned_conditioning([prompts] * batchsize)
+                z_enc = sampler.stochastic_encode(
+                    init_latent, torch.tensor([t_enc] * batchsize).to(device)
+                )
+                samples = sampler.decode(
+                    z_enc,
+                    c,
+                    t_enc,
+                    unconditional_guidance_scale=guidance_scale,
+                    unconditional_conditioning=uc,
+                )
+                # x_samples = latent_diffusion.decode_first_stage(samples) # Will result in Nan in output
+                # print(torch.sum(torch.isnan(samples)))
+                x_samples = latent_diffusion.decode_first_stage(samples)
+                # print(x_samples)
+                x_samples = latent_diffusion.decode_first_stage(samples[:,:,:-3,:])
+                # print(x_samples)
+                waveform = latent_diffusion.first_stage_model.decode_to_waveform(
+                    x_samples
+                )
+
+    return waveform
+
+def super_resolution_and_inpainting(
+    latent_diffusion,
+    text,
+    original_audio_file_path = None,
+    seed=42,
+    ddim_steps=200,
+    duration=None,
+    batchsize=1,
+    guidance_scale=2.5,
+    n_candidate_gen_per_text=3,
+    time_mask_ratio_start_and_end=(0.10, 0.15), # regenerate the 10% to 15% of the time steps in the spectrogram
+    # time_mask_ratio_start_and_end=(1.0, 1.0), # no inpainting
+    # freq_mask_ratio_start_and_end=(0.75, 1.0), # regenerate the higher 75% to 100% mel bins
+    freq_mask_ratio_start_and_end=(1.0, 1.0), # no super-resolution
+    config=None,
+):
+    seed_everything(int(seed))
+    if config is not None:
+        assert type(config) is str
+        config = yaml.load(open(config, "r"), Loader=yaml.FullLoader)
+    else:
+        config = default_audioldm_config()
+    fn_STFT = TacotronSTFT(
+        config["preprocessing"]["stft"]["filter_length"],
+        config["preprocessing"]["stft"]["hop_length"],
+        config["preprocessing"]["stft"]["win_length"],
+        config["preprocessing"]["mel"]["n_mel_channels"],
+        config["preprocessing"]["audio"]["sampling_rate"],
+        config["preprocessing"]["mel"]["mel_fmin"],
+        config["preprocessing"]["mel"]["mel_fmax"],
+    )
+    
+    # waveform = read_wav_file(original_audio_file_path, None)
+    mel, _, _ = wav_to_fbank(
+        original_audio_file_path, target_length=int(duration * 102.4), fn_STFT=fn_STFT
+    )
+    
+    batch = make_batch_for_text_to_audio(text, fbank=mel[None,...], batchsize=batchsize)
+        
+    # latent_diffusion.latent_t_size = duration_to_latent_t_size(duration)
+    latent_diffusion = set_cond_text(latent_diffusion)
+        
+    with torch.no_grad():
+        waveform = latent_diffusion.generate_sample_masked(
+            [batch],
+            unconditional_guidance_scale=guidance_scale,
+            ddim_steps=ddim_steps,
+            n_candidate_gen_per_text=n_candidate_gen_per_text,
+            duration=duration,
+            time_mask_ratio_start_and_end=time_mask_ratio_start_and_end,
+            freq_mask_ratio_start_and_end=freq_mask_ratio_start_and_end
+        )
+    return waveform

audioldm/utils.py

@@ -0,0 +1,281 @@
+import contextlib
+import importlib
+
+from inspect import isfunction
+import os
+import soundfile as sf
+import time
+import wave
+
+import urllib.request
+import progressbar
+
+CACHE_DIR = os.getenv(
+    "AUDIOLDM_CACHE_DIR",
+    os.path.join(os.path.expanduser("~"), ".cache/audioldm"))
+
+def get_duration(fname):
+    with contextlib.closing(wave.open(fname, 'r')) as f:
+        frames = f.getnframes()
+        rate = f.getframerate()
+        return frames / float(rate)
+    
+def get_bit_depth(fname):
+    with contextlib.closing(wave.open(fname, 'r')) as f:
+        bit_depth = f.getsampwidth() * 8
+        return bit_depth
+       
+def get_time():
+    t = time.localtime()
+    return time.strftime("%d_%m_%Y_%H_%M_%S", t)
+
+def seed_everything(seed):
+    import random, os
+    import numpy as np
+    import torch
+
+    random.seed(seed)
+    os.environ["PYTHONHASHSEED"] = str(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = True
+
+
+def save_wave(waveform, savepath, name="outwav"):
+    if type(name) is not list:
+        name = [name] * waveform.shape[0]
+
+    for i in range(waveform.shape[0]):
+        path = os.path.join(
+            savepath,
+            "%s_%s.wav"
+            % (
+                os.path.basename(name[i])
+                if (not ".wav" in name[i])
+                else os.path.basename(name[i]).split(".")[0],
+                i,
+            ),
+        )
+        print("Save audio to %s" % path)
+        sf.write(path, waveform[i, 0], samplerate=16000)
+
+
+def exists(x):
+    return x is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def count_params(model, verbose=False):
+    total_params = sum(p.numel() for p in model.parameters())
+    if verbose:
+        print(f"{model.__class__.__name__} has {total_params * 1.e-6:.2f} M params.")
+    return total_params
+
+
+def get_obj_from_str(string, reload=False):
+    module, cls = string.rsplit(".", 1)
+    if reload:
+        module_imp = importlib.import_module(module)
+        importlib.reload(module_imp)
+    return getattr(importlib.import_module(module, package=None), cls)
+
+
+def instantiate_from_config(config):
+    if not "target" in config:
+        if config == "__is_first_stage__":
+            return None
+        elif config == "__is_unconditional__":
+            return None
+        raise KeyError("Expected key `target` to instantiate.")
+    return get_obj_from_str(config["target"])(**config.get("params", dict()))
+
+
+def default_audioldm_config(model_name="audioldm-s-full"):    
+    basic_config = {
+        "wave_file_save_path": "./output",
+        "id": {
+            "version": "v1",
+            "name": "default",
+            "root": "/mnt/fast/nobackup/users/hl01486/projects/general_audio_generation/AudioLDM-python/config/default/latent_diffusion.yaml",
+        },
+        "preprocessing": {
+            "audio": {"sampling_rate": 16000, "max_wav_value": 32768},
+            "stft": {"filter_length": 1024, "hop_length": 160, "win_length": 1024},
+            "mel": {
+                "n_mel_channels": 64,
+                "mel_fmin": 0,
+                "mel_fmax": 8000,
+                "freqm": 0,
+                "timem": 0,
+                "blur": False,
+                "mean": -4.63,
+                "std": 2.74,
+                "target_length": 1024,
+            },
+        },
+        "model": {
+            "device": "cuda",
+            "target": "audioldm.pipline.LatentDiffusion",
+            "params": {
+                "base_learning_rate": 5e-06,
+                "linear_start": 0.0015,
+                "linear_end": 0.0195,
+                "num_timesteps_cond": 1,
+                "log_every_t": 200,
+                "timesteps": 1000,
+                "first_stage_key": "fbank",
+                "cond_stage_key": "waveform",
+                "latent_t_size": 256,
+                "latent_f_size": 16,
+                "channels": 8,
+                "cond_stage_trainable": True,
+                "conditioning_key": "film",
+                "monitor": "val/loss_simple_ema",
+                "scale_by_std": True,
+                "unet_config": {
+                    "target": "audioldm.latent_diffusion.openaimodel.UNetModel",
+                    "params": {
+                        "image_size": 64,
+                        "extra_film_condition_dim": 512,
+                        "extra_film_use_concat": True,
+                        "in_channels": 8,
+                        "out_channels": 8,
+                        "model_channels": 128,
+                        "attention_resolutions": [8, 4, 2],
+                        "num_res_blocks": 2,
+                        "channel_mult": [1, 2, 3, 5],
+                        "num_head_channels": 32,
+                        "use_spatial_transformer": True,
+                    },
+                },
+                "first_stage_config": {
+                    "base_learning_rate": 4.5e-05,
+                    "target": "audioldm.variational_autoencoder.autoencoder.AutoencoderKL",
+                    "params": {
+                        "monitor": "val/rec_loss",
+                        "image_key": "fbank",
+                        "subband": 1,
+                        "embed_dim": 8,
+                        "time_shuffle": 1,
+                        "ddconfig": {
+                            "double_z": True,
+                            "z_channels": 8,
+                            "resolution": 256,
+                            "downsample_time": False,
+                            "in_channels": 1,
+                            "out_ch": 1,
+                            "ch": 128,
+                            "ch_mult": [1, 2, 4],
+                            "num_res_blocks": 2,
+                            "attn_resolutions": [],
+                            "dropout": 0.0,
+                        },
+                    },
+                },
+                "cond_stage_config": {
+                    "target": "audioldm.clap.encoders.CLAPAudioEmbeddingClassifierFreev2",
+                    "params": {
+                        "key": "waveform",
+                        "sampling_rate": 16000,
+                        "embed_mode": "audio",
+                        "unconditional_prob": 0.1,
+                    },
+                },
+            },
+        },
+    }
+    
+    if("-l-" in model_name):
+        basic_config["model"]["params"]["unet_config"]["params"]["model_channels"] = 256
+        basic_config["model"]["params"]["unet_config"]["params"]["num_head_channels"] = 64
+    elif("-m-" in model_name):
+        basic_config["model"]["params"]["unet_config"]["params"]["model_channels"] = 192
+        basic_config["model"]["params"]["cond_stage_config"]["params"]["amodel"] = "HTSAT-base" # This model use a larger HTAST
+        
+    return basic_config
+        
+def get_metadata():
+    return {
+        "audioldm-s-full": {
+            "path": os.path.join(
+                CACHE_DIR,
+                "audioldm-s-full.ckpt",
+            ),
+            "url": "https://zenodo.org/record/7600541/files/audioldm-s-full?download=1",
+        },
+        "audioldm-l-full": {
+            "path": os.path.join(
+                CACHE_DIR,
+                "audioldm-l-full.ckpt",
+            ),
+            "url": "https://zenodo.org/record/7698295/files/audioldm-full-l.ckpt?download=1",
+        },
+        "audioldm-s-full-v2": {
+            "path": os.path.join(
+                CACHE_DIR,
+                "audioldm-s-full-v2.ckpt",
+            ),
+            "url": "https://zenodo.org/record/7698295/files/audioldm-full-s-v2.ckpt?download=1",
+        },
+        "audioldm-m-text-ft": {
+            "path": os.path.join(
+                CACHE_DIR,
+                "audioldm-m-text-ft.ckpt",
+            ),
+            "url": "https://zenodo.org/record/7813012/files/audioldm-m-text-ft.ckpt?download=1",
+        },
+        "audioldm-s-text-ft": {
+            "path": os.path.join(
+                CACHE_DIR,
+                "audioldm-s-text-ft.ckpt",
+            ),
+            "url": "https://zenodo.org/record/7813012/files/audioldm-s-text-ft.ckpt?download=1",
+        },
+        "audioldm-m-full": {
+            "path": os.path.join(
+                CACHE_DIR,
+                "audioldm-m-full.ckpt",
+            ),
+            "url": "https://zenodo.org/record/7813012/files/audioldm-m-full.ckpt?download=1",
+        },
+    }
+    
+class MyProgressBar():
+    def __init__(self):
+        self.pbar = None
+
+    def __call__(self, block_num, block_size, total_size):
+        if not self.pbar:
+            self.pbar=progressbar.ProgressBar(maxval=total_size)
+            self.pbar.start()
+
+        downloaded = block_num * block_size
+        if downloaded < total_size:
+            self.pbar.update(downloaded)
+        else:
+            self.pbar.finish()
+            
+def download_checkpoint(checkpoint_name="audioldm-s-full"):
+    meta = get_metadata()
+    if(checkpoint_name not in meta.keys()):
+        print("The model name you provided is not supported. Please use one of the following: ", meta.keys())
+
+    if not os.path.exists(meta[checkpoint_name]["path"]) or os.path.getsize(meta[checkpoint_name]["path"]) < 2*10**9:
+        os.makedirs(os.path.dirname(meta[checkpoint_name]["path"]), exist_ok=True)
+        print(f"Downloading the main structure of {checkpoint_name} into {os.path.dirname(meta[checkpoint_name]['path'])}")
+
+        urllib.request.urlretrieve(meta[checkpoint_name]["url"], meta[checkpoint_name]["path"], MyProgressBar())
+        print(
+            "Weights downloaded in: {} Size: {}".format(
+                meta[checkpoint_name]["path"],
+                os.path.getsize(meta[checkpoint_name]["path"]),
+            )
+        )
+    

audioldm/variational_autoencoder/__init__.py

@@ -0,0 +1 @@
+from .autoencoder import AutoencoderKL

audioldm/variational_autoencoder/autoencoder.py

@@ -0,0 +1,135 @@
+import torch
+from audioldm.latent_diffusion.ema import *
+from audioldm.variational_autoencoder.modules import Encoder, Decoder
+from audioldm.variational_autoencoder.distributions import DiagonalGaussianDistribution
+
+from audioldm.hifigan.utilities import get_vocoder, vocoder_infer
+
+
+class AutoencoderKL(nn.Module):
+    def __init__(
+        self,
+        ddconfig=None,
+        lossconfig=None,
+        image_key="fbank",
+        embed_dim=None,
+        time_shuffle=1,
+        subband=1,
+        ckpt_path=None,
+        reload_from_ckpt=None,
+        ignore_keys=[],
+        colorize_nlabels=None,
+        monitor=None,
+        base_learning_rate=1e-5,
+        scale_factor=1
+    ):
+        super().__init__()
+
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+
+        self.subband = int(subband)
+
+        if self.subband > 1:
+            print("Use subband decomposition %s" % self.subband)
+
+        self.quant_conv = torch.nn.Conv2d(2 * ddconfig["z_channels"], 2 * embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+
+        self.vocoder = get_vocoder(None, "cpu")
+        self.embed_dim = embed_dim
+
+        if monitor is not None:
+            self.monitor = monitor
+
+        self.time_shuffle = time_shuffle
+        self.reload_from_ckpt = reload_from_ckpt
+        self.reloaded = False
+        self.mean, self.std = None, None
+        
+        self.scale_factor = scale_factor
+
+    def encode(self, x):
+        # x = self.time_shuffle_operation(x)
+        x = self.freq_split_subband(x)
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+        dec = self.freq_merge_subband(dec)
+        return dec
+
+    def decode_to_waveform(self, dec):
+        dec = dec.squeeze(1).permute(0, 2, 1)
+        wav_reconstruction = vocoder_infer(dec, self.vocoder)
+        return wav_reconstruction
+
+    def forward(self, input, sample_posterior=True):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+
+        if self.flag_first_run:
+            print("Latent size: ", z.size())
+            self.flag_first_run = False
+
+        dec = self.decode(z)
+
+        return dec, posterior
+
+    def freq_split_subband(self, fbank):
+        if self.subband == 1 or self.image_key != "stft":
+            return fbank
+
+        bs, ch, tstep, fbins = fbank.size()
+
+        assert fbank.size(-1) % self.subband == 0
+        assert ch == 1
+
+        return (
+            fbank.squeeze(1)
+            .reshape(bs, tstep, self.subband, fbins // self.subband)
+            .permute(0, 2, 1, 3)
+        )
+
+    def freq_merge_subband(self, subband_fbank):
+        if self.subband == 1 or self.image_key != "stft":
+            return subband_fbank
+        assert subband_fbank.size(1) == self.subband  # Channel dimension
+        bs, sub_ch, tstep, fbins = subband_fbank.size()
+        return subband_fbank.permute(0, 2, 1, 3).reshape(bs, tstep, -1).unsqueeze(1)
+    
+    def device(self):
+        return next(self.parameters()).device
+    
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        return self.encode(x)
+    
+    @torch.no_grad()
+    def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, "b h w c -> b c h w").contiguous()
+
+        z = 1.0 / self.scale_factor * z
+        return self.decode(z)
+
+    def get_first_stage_encoding(self, encoder_posterior):
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample()
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(
+                f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented"
+            )
+        return self.scale_factor * z

audioldm/variational_autoencoder/distributions.py

@@ -0,0 +1,102 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(
+                device=self.parameters.device
+            )
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape).to(
+            device=self.parameters.device
+        )
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.0])
+        else:
+            if other is None:
+                return 0.5 * torch.mean(
+                    torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar,
+                    dim=[1, 2, 3],
+                )
+            else:
+                return 0.5 * torch.mean(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var
+                    - 1.0
+                    - self.logvar
+                    + other.logvar,
+                    dim=[1, 2, 3],
+                )
+
+    def nll(self, sample, dims=[1, 2, 3]):
+        if self.deterministic:
+            return torch.Tensor([0.0])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims,
+        )
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

audioldm/variational_autoencoder/modules.py

@@ -0,0 +1,1066 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+
+from audioldm.utils import instantiate_from_config
+from audioldm.latent_diffusion.attention import LinearAttention
+
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return x * torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(
+        num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True
+    )
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=3, stride=1, padding=1
+            )
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class UpsampleTimeStride4(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=5, stride=1, padding=2
+            )
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=(4.0, 2.0), mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # Do time downsampling here
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=3, stride=2, padding=0
+            )
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0, 1, 0, 1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class DownsampleTimeStride4(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # Do time downsampling here
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=5, stride=(4, 2), padding=1
+            )
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0, 1, 0, 1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=(4, 2), stride=(4, 2))
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout,
+        temb_channels=512,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(
+                    in_channels, out_channels, kernel_size=3, stride=1, padding=1
+                )
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(
+                    in_channels, out_channels, kernel_size=1, stride=1, padding=0
+                )
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class LinAttnBlock(LinearAttention):
+    """to match AttnBlock usage"""
+
+    def __init__(self, in_channels):
+        super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = q.reshape(b, c, h * w).contiguous()
+        q = q.permute(0, 2, 1).contiguous()  # b,hw,c
+        k = k.reshape(b, c, h * w).contiguous()  # b,c,hw
+        w_ = torch.bmm(q, k).contiguous()  # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b, c, h * w).contiguous()
+        w_ = w_.permute(0, 2, 1).contiguous()  # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(
+            v, w_
+        ).contiguous()  # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b, c, h, w).contiguous()
+
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+def make_attn(in_channels, attn_type="vanilla"):
+    assert attn_type in ["vanilla", "linear", "none"], f"attn_type {attn_type} unknown"
+    # print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        return AttnBlock(in_channels)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        return LinAttnBlock(in_channels)
+
+
+class Model(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        use_timestep=True,
+        use_linear_attn=False,
+        attn_type="vanilla",
+    ):
+        super().__init__()
+        if use_linear_attn:
+            attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch * 4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList(
+                [
+                    torch.nn.Linear(self.ch, self.temb_ch),
+                    torch.nn.Linear(self.temb_ch, self.temb_ch),
+                ]
+            )
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(
+            in_channels, self.ch, kernel_size=3, stride=1, padding=1
+        )
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            skip_in = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch * in_ch_mult[i_level]
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in + skip_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in, out_ch, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x, t=None, context=None):
+        # assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb
+                )
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        double_z=True,
+        use_linear_attn=False,
+        attn_type="vanilla",
+        downsample_time_stride4_levels=[],
+        **ignore_kwargs,
+    ):
+        super().__init__()
+        if use_linear_attn:
+            attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.downsample_time_stride4_levels = downsample_time_stride4_levels
+
+        if len(self.downsample_time_stride4_levels) > 0:
+            assert max(self.downsample_time_stride4_levels) < self.num_resolutions, (
+                "The level to perform downsample 4 operation need to be smaller than the total resolution number %s"
+                % str(self.num_resolutions)
+            )
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(
+            in_channels, self.ch, kernel_size=3, stride=1, padding=1
+        )
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                if i_level in self.downsample_time_stride4_levels:
+                    down.downsample = DownsampleTimeStride4(block_in, resamp_with_conv)
+                else:
+                    down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in,
+            2 * z_channels if double_z else z_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        give_pre_end=False,
+        tanh_out=False,
+        use_linear_attn=False,
+        downsample_time_stride4_levels=[],
+        attn_type="vanilla",
+        **ignorekwargs,
+    ):
+        super().__init__()
+        if use_linear_attn:
+            attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+        self.downsample_time_stride4_levels = downsample_time_stride4_levels
+
+        if len(self.downsample_time_stride4_levels) > 0:
+            assert max(self.downsample_time_stride4_levels) < self.num_resolutions, (
+                "The level to perform downsample 4 operation need to be smaller than the total resolution number %s"
+                % str(self.num_resolutions)
+            )
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,) + tuple(ch_mult)
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+        # print("Working with z of shape {} = {} dimensions.".format(
+        # self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(
+            z_channels, block_in, kernel_size=3, stride=1, padding=1
+        )
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                if i_level - 1 in self.downsample_time_stride4_levels:
+                    up.upsample = UpsampleTimeStride4(block_in, resamp_with_conv)
+                else:
+                    up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in, out_ch, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, z):
+        # assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class SimpleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, *args, **kwargs):
+        super().__init__()
+        self.model = nn.ModuleList(
+            [
+                nn.Conv2d(in_channels, in_channels, 1),
+                ResnetBlock(
+                    in_channels=in_channels,
+                    out_channels=2 * in_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                ),
+                ResnetBlock(
+                    in_channels=2 * in_channels,
+                    out_channels=4 * in_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                ),
+                ResnetBlock(
+                    in_channels=4 * in_channels,
+                    out_channels=2 * in_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                ),
+                nn.Conv2d(2 * in_channels, in_channels, 1),
+                Upsample(in_channels, with_conv=True),
+            ]
+        )
+        # end
+        self.norm_out = Normalize(in_channels)
+        self.conv_out = torch.nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x):
+        for i, layer in enumerate(self.model):
+            if i in [1, 2, 3]:
+                x = layer(x, None)
+            else:
+                x = layer(x)
+
+        h = self.norm_out(x)
+        h = nonlinearity(h)
+        x = self.conv_out(h)
+        return x
+
+
+class UpsampleDecoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        ch,
+        num_res_blocks,
+        resolution,
+        ch_mult=(2, 2),
+        dropout=0.0,
+    ):
+        super().__init__()
+        # upsampling
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        block_in = in_channels
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.res_blocks = nn.ModuleList()
+        self.upsample_blocks = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            res_block = []
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                res_block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+            self.res_blocks.append(nn.ModuleList(res_block))
+            if i_level != self.num_resolutions - 1:
+                self.upsample_blocks.append(Upsample(block_in, True))
+                curr_res = curr_res * 2
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x):
+        # upsampling
+        h = x
+        for k, i_level in enumerate(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.res_blocks[i_level][i_block](h, None)
+            if i_level != self.num_resolutions - 1:
+                h = self.upsample_blocks[k](h)
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class LatentRescaler(nn.Module):
+    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
+        super().__init__()
+        # residual block, interpolate, residual block
+        self.factor = factor
+        self.conv_in = nn.Conv2d(
+            in_channels, mid_channels, kernel_size=3, stride=1, padding=1
+        )
+        self.res_block1 = nn.ModuleList(
+            [
+                ResnetBlock(
+                    in_channels=mid_channels,
+                    out_channels=mid_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.attn = AttnBlock(mid_channels)
+        self.res_block2 = nn.ModuleList(
+            [
+                ResnetBlock(
+                    in_channels=mid_channels,
+                    out_channels=mid_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+        self.conv_out = nn.Conv2d(
+            mid_channels,
+            out_channels,
+            kernel_size=1,
+        )
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        for block in self.res_block1:
+            x = block(x, None)
+        x = torch.nn.functional.interpolate(
+            x,
+            size=(
+                int(round(x.shape[2] * self.factor)),
+                int(round(x.shape[3] * self.factor)),
+            ),
+        )
+        x = self.attn(x).contiguous()
+        for block in self.res_block2:
+            x = block(x, None)
+        x = self.conv_out(x)
+        return x
+
+
+class MergedRescaleEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        ch,
+        resolution,
+        out_ch,
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        ch_mult=(1, 2, 4, 8),
+        rescale_factor=1.0,
+        rescale_module_depth=1,
+    ):
+        super().__init__()
+        intermediate_chn = ch * ch_mult[-1]
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            num_res_blocks=num_res_blocks,
+            ch=ch,
+            ch_mult=ch_mult,
+            z_channels=intermediate_chn,
+            double_z=False,
+            resolution=resolution,
+            attn_resolutions=attn_resolutions,
+            dropout=dropout,
+            resamp_with_conv=resamp_with_conv,
+            out_ch=None,
+        )
+        self.rescaler = LatentRescaler(
+            factor=rescale_factor,
+            in_channels=intermediate_chn,
+            mid_channels=intermediate_chn,
+            out_channels=out_ch,
+            depth=rescale_module_depth,
+        )
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.rescaler(x)
+        return x
+
+
+class MergedRescaleDecoder(nn.Module):
+    def __init__(
+        self,
+        z_channels,
+        out_ch,
+        resolution,
+        num_res_blocks,
+        attn_resolutions,
+        ch,
+        ch_mult=(1, 2, 4, 8),
+        dropout=0.0,
+        resamp_with_conv=True,
+        rescale_factor=1.0,
+        rescale_module_depth=1,
+    ):
+        super().__init__()
+        tmp_chn = z_channels * ch_mult[-1]
+        self.decoder = Decoder(
+            out_ch=out_ch,
+            z_channels=tmp_chn,
+            attn_resolutions=attn_resolutions,
+            dropout=dropout,
+            resamp_with_conv=resamp_with_conv,
+            in_channels=None,
+            num_res_blocks=num_res_blocks,
+            ch_mult=ch_mult,
+            resolution=resolution,
+            ch=ch,
+        )
+        self.rescaler = LatentRescaler(
+            factor=rescale_factor,
+            in_channels=z_channels,
+            mid_channels=tmp_chn,
+            out_channels=tmp_chn,
+            depth=rescale_module_depth,
+        )
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Upsampler(nn.Module):
+    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
+        super().__init__()
+        assert out_size >= in_size
+        num_blocks = int(np.log2(out_size // in_size)) + 1
+        factor_up = 1.0 + (out_size % in_size)
+        print(
+            f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}"
+        )
+        self.rescaler = LatentRescaler(
+            factor=factor_up,
+            in_channels=in_channels,
+            mid_channels=2 * in_channels,
+            out_channels=in_channels,
+        )
+        self.decoder = Decoder(
+            out_ch=out_channels,
+            resolution=out_size,
+            z_channels=in_channels,
+            num_res_blocks=2,
+            attn_resolutions=[],
+            in_channels=None,
+            ch=in_channels,
+            ch_mult=[ch_mult for _ in range(num_blocks)],
+        )
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Resize(nn.Module):
+    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
+        super().__init__()
+        self.with_conv = learned
+        self.mode = mode
+        if self.with_conv:
+            print(
+                f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode"
+            )
+            raise NotImplementedError()
+            assert in_channels is not None
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=4, stride=2, padding=1
+            )
+
+    def forward(self, x, scale_factor=1.0):
+        if scale_factor == 1.0:
+            return x
+        else:
+            x = torch.nn.functional.interpolate(
+                x, mode=self.mode, align_corners=False, scale_factor=scale_factor
+            )
+        return x
+
+
+class FirstStagePostProcessor(nn.Module):
+    def __init__(
+        self,
+        ch_mult: list,
+        in_channels,
+        pretrained_model: nn.Module = None,
+        reshape=False,
+        n_channels=None,
+        dropout=0.0,
+        pretrained_config=None,
+    ):
+        super().__init__()
+        if pretrained_config is None:
+            assert (
+                pretrained_model is not None
+            ), 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.pretrained_model = pretrained_model
+        else:
+            assert (
+                pretrained_config is not None
+            ), 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.instantiate_pretrained(pretrained_config)
+
+        self.do_reshape = reshape
+
+        if n_channels is None:
+            n_channels = self.pretrained_model.encoder.ch
+
+        self.proj_norm = Normalize(in_channels, num_groups=in_channels // 2)
+        self.proj = nn.Conv2d(
+            in_channels, n_channels, kernel_size=3, stride=1, padding=1
+        )
+
+        blocks = []
+        downs = []
+        ch_in = n_channels
+        for m in ch_mult:
+            blocks.append(
+                ResnetBlock(
+                    in_channels=ch_in, out_channels=m * n_channels, dropout=dropout
+                )
+            )
+            ch_in = m * n_channels
+            downs.append(Downsample(ch_in, with_conv=False))
+
+        self.model = nn.ModuleList(blocks)
+        self.downsampler = nn.ModuleList(downs)
+
+    def instantiate_pretrained(self, config):
+        model = instantiate_from_config(config)
+        self.pretrained_model = model.eval()
+        # self.pretrained_model.train = False
+        for param in self.pretrained_model.parameters():
+            param.requires_grad = False
+
+    @torch.no_grad()
+    def encode_with_pretrained(self, x):
+        c = self.pretrained_model.encode(x)
+        if isinstance(c, DiagonalGaussianDistribution):
+            c = c.mode()
+        return c
+
+    def forward(self, x):
+        z_fs = self.encode_with_pretrained(x)
+        z = self.proj_norm(z_fs)
+        z = self.proj(z)
+        z = nonlinearity(z)
+
+        for submodel, downmodel in zip(self.model, self.downsampler):
+            z = submodel(z, temb=None)
+            z = downmodel(z)
+
+        if self.do_reshape:
+            z = rearrange(z, "b c h w -> b (h w) c")
+        return z

diffusers/CITATION.cff

@@ -0,0 +1,40 @@
+cff-version: 1.2.0
+title: 'Diffusers: State-of-the-art diffusion models'
+message: >-
+  If you use this software, please cite it using the
+  metadata from this file.
+type: software
+authors:
+  - given-names: Patrick
+    family-names: von Platen
+  - given-names: Suraj
+    family-names: Patil
+  - given-names: Anton
+    family-names: Lozhkov
+  - given-names: Pedro
+    family-names: Cuenca
+  - given-names: Nathan
+    family-names: Lambert
+  - given-names: Kashif
+    family-names: Rasul
+  - given-names: Mishig
+    family-names: Davaadorj
+  - given-names: Thomas
+    family-names: Wolf
+repository-code: 'https://github.com/huggingface/diffusers'
+abstract: >-
+  Diffusers provides pretrained diffusion models across
+  multiple modalities, such as vision and audio, and serves
+  as a modular toolbox for inference and training of
+  diffusion models.
+keywords:
+  - deep-learning
+  - pytorch
+  - image-generation
+  - diffusion
+  - text2image
+  - image2image
+  - score-based-generative-modeling
+  - stable-diffusion
+license: Apache-2.0
+version: 0.12.1

diffusers/CODE_OF_CONDUCT.md

@@ -0,0 +1,130 @@
+
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+We as members, contributors, and leaders pledge to make participation in our
+community a harassment-free experience for everyone, regardless of age, body
+size, visible or invisible disability, ethnicity, sex characteristics, gender
+identity and expression, level of experience, education, socio-economic status,
+nationality, personal appearance, race, religion, or sexual identity
+and orientation.
+
+We pledge to act and interact in ways that contribute to an open, welcoming,
+diverse, inclusive, and healthy community.
+
+## Our Standards
+
+Examples of behavior that contributes to a positive environment for our
+community include:
+
+* Demonstrating empathy and kindness toward other people
+* Being respectful of differing opinions, viewpoints, and experiences
+* Giving and gracefully accepting constructive feedback
+* Accepting responsibility and apologizing to those affected by our mistakes,
+  and learning from the experience
+* Focusing on what is best not just for us as individuals, but for the
+  overall diffusers community
+
+Examples of unacceptable behavior include:
+
+* The use of sexualized language or imagery, and sexual attention or
+  advances of any kind
+* Trolling, insulting or derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or email
+  address, without their explicit permission
+* Spamming issues or PRs with links to projects unrelated to this library
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Enforcement Responsibilities
+
+Community leaders are responsible for clarifying and enforcing our standards of
+acceptable behavior and will take appropriate and fair corrective action in
+response to any behavior that they deem inappropriate, threatening, offensive,
+or harmful.
+
+Community leaders have the right and responsibility to remove, edit, or reject
+comments, commits, code, wiki edits, issues, and other contributions that are
+not aligned to this Code of Conduct, and will communicate reasons for moderation
+decisions when appropriate.
+
+## Scope
+
+This Code of Conduct applies within all community spaces, and also applies when
+an individual is officially representing the community in public spaces.
+Examples of representing our community include using an official e-mail address,
+posting via an official social media account, or acting as an appointed
+representative at an online or offline event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported to the community leaders responsible for enforcement at
+feedback@huggingface.co.
+All complaints will be reviewed and investigated promptly and fairly.
+
+All community leaders are obligated to respect the privacy and security of the
+reporter of any incident.
+
+## Enforcement Guidelines
+
+Community leaders will follow these Community Impact Guidelines in determining
+the consequences for any action they deem in violation of this Code of Conduct:
+
+### 1. Correction
+
+**Community Impact**: Use of inappropriate language or other behavior deemed
+unprofessional or unwelcome in the community.
+
+**Consequence**: A private, written warning from community leaders, providing
+clarity around the nature of the violation and an explanation of why the
+behavior was inappropriate. A public apology may be requested.
+
+### 2. Warning
+
+**Community Impact**: A violation through a single incident or series
+of actions.
+
+**Consequence**: A warning with consequences for continued behavior. No
+interaction with the people involved, including unsolicited interaction with
+those enforcing the Code of Conduct, for a specified period of time. This
+includes avoiding interactions in community spaces as well as external channels
+like social media. Violating these terms may lead to a temporary or
+permanent ban.
+
+### 3. Temporary Ban
+
+**Community Impact**: A serious violation of community standards, including
+sustained inappropriate behavior.
+
+**Consequence**: A temporary ban from any sort of interaction or public
+communication with the community for a specified period of time. No public or
+private interaction with the people involved, including unsolicited interaction
+with those enforcing the Code of Conduct, is allowed during this period.
+Violating these terms may lead to a permanent ban.
+
+### 4. Permanent Ban
+
+**Community Impact**: Demonstrating a pattern of violation of community
+standards, including sustained inappropriate behavior,  harassment of an
+individual, or aggression toward or disparagement of classes of individuals.
+
+**Consequence**: A permanent ban from any sort of public interaction within
+the community.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage],
+version 2.0, available at
+https://www.contributor-covenant.org/version/2/0/code_of_conduct.html.
+
+Community Impact Guidelines were inspired by [Mozilla's code of conduct
+enforcement ladder](https://github.com/mozilla/diversity).
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see the FAQ at
+https://www.contributor-covenant.org/faq. Translations are available at
+https://www.contributor-covenant.org/translations.