Upload 11 files
Browse files- SD/attention.py +122 -0
- SD/clip.py +96 -0
- SD/ddpm.py +123 -0
- SD/decoder.py +177 -0
- SD/diffusion.py +349 -0
- SD/encoder.py +103 -0
- SD/model_converter.py +0 -0
- SD/model_loader.py +28 -0
- SD/pipeline.py +170 -0
- SD/run.py +64 -0
- SD/sd_demo.ipynb +0 -0
SD/attention.py
ADDED
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import math
|
5 |
+
|
6 |
+
class SelfAttention(nn.Module):
|
7 |
+
def __init__(self, n_heads, d_embed, in_proj_bias=True, out_proj_bias=True):
|
8 |
+
super().__init__()
|
9 |
+
# This combines the Wq, Wk and Wv matrices into one matrix
|
10 |
+
self.in_proj = nn.Linear(d_embed, 3 * d_embed, bias=in_proj_bias)
|
11 |
+
# This one represents the Wo matrix
|
12 |
+
self.out_proj = nn.Linear(d_embed, d_embed, bias=out_proj_bias)
|
13 |
+
self.n_heads = n_heads
|
14 |
+
self.d_head = d_embed // n_heads
|
15 |
+
|
16 |
+
def forward(self, x, causal_mask=False):
|
17 |
+
# x: # (Batch_Size, Seq_Len, Dim)
|
18 |
+
|
19 |
+
# (Batch_Size, Seq_Len, Dim)
|
20 |
+
input_shape = x.shape
|
21 |
+
|
22 |
+
# (Batch_Size, Seq_Len, Dim)
|
23 |
+
batch_size, sequence_length, d_embed = input_shape
|
24 |
+
|
25 |
+
# (Batch_Size, Seq_Len, H, Dim / H)
|
26 |
+
interim_shape = (batch_size, sequence_length, self.n_heads, self.d_head)
|
27 |
+
|
28 |
+
# (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim * 3) -> 3 tensor of shape (Batch_Size, Seq_Len, Dim)
|
29 |
+
q, k, v = self.in_proj(x).chunk(3, dim=-1)
|
30 |
+
|
31 |
+
# (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, H, Dim / H) -> (Batch_Size, H, Seq_Len, Dim / H)
|
32 |
+
q = q.view(interim_shape).transpose(1, 2)
|
33 |
+
k = k.view(interim_shape).transpose(1, 2)
|
34 |
+
v = v.view(interim_shape).transpose(1, 2)
|
35 |
+
|
36 |
+
# (Batch_Size, H, Seq_Len, Dim) @ (Batch_Size, H, Dim, Seq_Len) -> (Batch_Size, H, Seq_Len, Seq_Len)
|
37 |
+
weight = q @ k.transpose(-1, -2)
|
38 |
+
|
39 |
+
if causal_mask:
|
40 |
+
# Mask where the upper triangle (above the principal diagonal) is 1
|
41 |
+
mask = torch.ones_like(weight, dtype=torch.bool).triu(1)
|
42 |
+
# Fill the upper triangle with -inf
|
43 |
+
weight.masked_fill_(mask, -torch.inf)
|
44 |
+
|
45 |
+
# Divide by d_k (Dim / H).
|
46 |
+
# (Batch_Size, H, Seq_Len, Seq_Len) -> (Batch_Size, H, Seq_Len, Seq_Len)
|
47 |
+
weight /= math.sqrt(self.d_head)
|
48 |
+
|
49 |
+
# (Batch_Size, H, Seq_Len, Seq_Len) -> (Batch_Size, H, Seq_Len, Seq_Len)
|
50 |
+
weight = F.softmax(weight, dim=-1)
|
51 |
+
|
52 |
+
# (Batch_Size, H, Seq_Len, Seq_Len) @ (Batch_Size, H, Seq_Len, Dim / H) -> (Batch_Size, H, Seq_Len, Dim / H)
|
53 |
+
output = weight @ v
|
54 |
+
|
55 |
+
# (Batch_Size, H, Seq_Len, Dim / H) -> (Batch_Size, Seq_Len, H, Dim / H)
|
56 |
+
output = output.transpose(1, 2)
|
57 |
+
|
58 |
+
# (Batch_Size, Seq_Len, H, Dim / H) -> (Batch_Size, Seq_Len, Dim)
|
59 |
+
output = output.reshape(input_shape)
|
60 |
+
|
61 |
+
# (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
|
62 |
+
output = self.out_proj(output)
|
63 |
+
|
64 |
+
# (Batch_Size, Seq_Len, Dim)
|
65 |
+
return output
|
66 |
+
|
67 |
+
class CrossAttention(nn.Module):
|
68 |
+
def __init__(self, n_heads, d_embed, d_cross, in_proj_bias=True, out_proj_bias=True):
|
69 |
+
super().__init__()
|
70 |
+
self.q_proj = nn.Linear(d_embed, d_embed, bias=in_proj_bias)
|
71 |
+
self.k_proj = nn.Linear(d_cross, d_embed, bias=in_proj_bias)
|
72 |
+
self.v_proj = nn.Linear(d_cross, d_embed, bias=in_proj_bias)
|
73 |
+
self.out_proj = nn.Linear(d_embed, d_embed, bias=out_proj_bias)
|
74 |
+
self.n_heads = n_heads
|
75 |
+
self.d_head = d_embed // n_heads
|
76 |
+
|
77 |
+
def forward(self, x, y):
|
78 |
+
# x (latent): # (Batch_Size, Seq_Len_Q, Dim_Q)
|
79 |
+
# y (context): # (Batch_Size, Seq_Len_KV, Dim_KV) = (Batch_Size, 77, 768)
|
80 |
+
|
81 |
+
input_shape = x.shape
|
82 |
+
batch_size, sequence_length, d_embed = input_shape
|
83 |
+
# Divide each embedding of Q into multiple heads such that d_heads * n_heads = Dim_Q
|
84 |
+
interim_shape = (batch_size, -1, self.n_heads, self.d_head)
|
85 |
+
|
86 |
+
# (Batch_Size, Seq_Len_Q, Dim_Q) -> (Batch_Size, Seq_Len_Q, Dim_Q)
|
87 |
+
q = self.q_proj(x)
|
88 |
+
# (Batch_Size, Seq_Len_KV, Dim_KV) -> (Batch_Size, Seq_Len_KV, Dim_Q)
|
89 |
+
k = self.k_proj(y)
|
90 |
+
# (Batch_Size, Seq_Len_KV, Dim_KV) -> (Batch_Size, Seq_Len_KV, Dim_Q)
|
91 |
+
v = self.v_proj(y)
|
92 |
+
|
93 |
+
# (Batch_Size, Seq_Len_Q, Dim_Q) -> (Batch_Size, Seq_Len_Q, H, Dim_Q / H) -> (Batch_Size, H, Seq_Len_Q, Dim_Q / H)
|
94 |
+
q = q.view(interim_shape).transpose(1, 2)
|
95 |
+
# (Batch_Size, Seq_Len_KV, Dim_Q) -> (Batch_Size, Seq_Len_KV, H, Dim_Q / H) -> (Batch_Size, H, Seq_Len_KV, Dim_Q / H)
|
96 |
+
k = k.view(interim_shape).transpose(1, 2)
|
97 |
+
# (Batch_Size, Seq_Len_KV, Dim_Q) -> (Batch_Size, Seq_Len_KV, H, Dim_Q / H) -> (Batch_Size, H, Seq_Len_KV, Dim_Q / H)
|
98 |
+
v = v.view(interim_shape).transpose(1, 2)
|
99 |
+
|
100 |
+
# (Batch_Size, H, Seq_Len_Q, Dim_Q / H) @ (Batch_Size, H, Dim_Q / H, Seq_Len_KV) -> (Batch_Size, H, Seq_Len_Q, Seq_Len_KV)
|
101 |
+
weight = q @ k.transpose(-1, -2)
|
102 |
+
|
103 |
+
# (Batch_Size, H, Seq_Len_Q, Seq_Len_KV)
|
104 |
+
weight /= math.sqrt(self.d_head)
|
105 |
+
|
106 |
+
# (Batch_Size, H, Seq_Len_Q, Seq_Len_KV)
|
107 |
+
weight = F.softmax(weight, dim=-1)
|
108 |
+
|
109 |
+
# (Batch_Size, H, Seq_Len_Q, Seq_Len_KV) @ (Batch_Size, H, Seq_Len_KV, Dim_Q / H) -> (Batch_Size, H, Seq_Len_Q, Dim_Q / H)
|
110 |
+
output = weight @ v
|
111 |
+
|
112 |
+
# (Batch_Size, H, Seq_Len_Q, Dim_Q / H) -> (Batch_Size, Seq_Len_Q, H, Dim_Q / H)
|
113 |
+
output = output.transpose(1, 2).contiguous()
|
114 |
+
|
115 |
+
# (Batch_Size, Seq_Len_Q, H, Dim_Q / H) -> (Batch_Size, Seq_Len_Q, Dim_Q)
|
116 |
+
output = output.view(input_shape)
|
117 |
+
|
118 |
+
# (Batch_Size, Seq_Len_Q, Dim_Q) -> (Batch_Size, Seq_Len_Q, Dim_Q)
|
119 |
+
output = self.out_proj(output)
|
120 |
+
|
121 |
+
# (Batch_Size, Seq_Len_Q, Dim_Q)
|
122 |
+
return output
|
SD/clip.py
ADDED
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from attention import SelfAttention
|
5 |
+
|
6 |
+
class CLIPEmbedding(nn.Module):
|
7 |
+
def __init__(self, n_vocab: int, n_embd: int, n_token: int):
|
8 |
+
super().__init__()
|
9 |
+
|
10 |
+
self.token_embedding = nn.Embedding(n_vocab, n_embd)
|
11 |
+
# A learnable weight matrix encodes the position information for each token
|
12 |
+
self.position_embedding = nn.Parameter(torch.zeros((n_token, n_embd)))
|
13 |
+
|
14 |
+
def forward(self, tokens):
|
15 |
+
# (Batch_Size, Seq_Len) -> (Batch_Size, Seq_Len, Dim)
|
16 |
+
x = self.token_embedding(tokens)
|
17 |
+
# (Batch_Size, Seq_Len) -> (Batch_Size, Seq_Len, Dim)
|
18 |
+
x += self.position_embedding
|
19 |
+
|
20 |
+
return x
|
21 |
+
|
22 |
+
class CLIPLayer(nn.Module):
|
23 |
+
def __init__(self, n_head: int, n_embd: int):
|
24 |
+
super().__init__()
|
25 |
+
|
26 |
+
# Pre-attention norm
|
27 |
+
self.layernorm_1 = nn.LayerNorm(n_embd)
|
28 |
+
# Self attention
|
29 |
+
self.attention = SelfAttention(n_head, n_embd)
|
30 |
+
# Pre-FNN norm
|
31 |
+
self.layernorm_2 = nn.LayerNorm(n_embd)
|
32 |
+
# Feedforward layer
|
33 |
+
self.linear_1 = nn.Linear(n_embd, 4 * n_embd)
|
34 |
+
self.linear_2 = nn.Linear(4 * n_embd, n_embd)
|
35 |
+
|
36 |
+
def forward(self, x):
|
37 |
+
# (Batch_Size, Seq_Len, Dim)
|
38 |
+
residue = x
|
39 |
+
|
40 |
+
### SELF ATTENTION ###
|
41 |
+
|
42 |
+
# (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
|
43 |
+
x = self.layernorm_1(x)
|
44 |
+
|
45 |
+
# (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
|
46 |
+
x = self.attention(x, causal_mask=True)
|
47 |
+
|
48 |
+
# (Batch_Size, Seq_Len, Dim) + (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
|
49 |
+
x += residue
|
50 |
+
|
51 |
+
### FEEDFORWARD LAYER ###
|
52 |
+
# Apply a feedforward layer where the hidden dimension is 4 times the embedding dimension.
|
53 |
+
|
54 |
+
residue = x
|
55 |
+
# (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
|
56 |
+
x = self.layernorm_2(x)
|
57 |
+
|
58 |
+
# (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, 4 * Dim)
|
59 |
+
x = self.linear_1(x)
|
60 |
+
|
61 |
+
# (Batch_Size, Seq_Len, 4 * Dim) -> (Batch_Size, Seq_Len, 4 * Dim)
|
62 |
+
x = x * torch.sigmoid(1.702 * x) # QuickGELU activation function
|
63 |
+
|
64 |
+
# (Batch_Size, Seq_Len, 4 * Dim) -> (Batch_Size, Seq_Len, Dim)
|
65 |
+
x = self.linear_2(x)
|
66 |
+
|
67 |
+
# (Batch_Size, Seq_Len, Dim) + (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
|
68 |
+
x += residue
|
69 |
+
|
70 |
+
return x
|
71 |
+
|
72 |
+
class CLIP(nn.Module):
|
73 |
+
def __init__(self):
|
74 |
+
super().__init__()
|
75 |
+
self.embedding = CLIPEmbedding(49408, 768, 77)
|
76 |
+
|
77 |
+
self.layers = nn.ModuleList([
|
78 |
+
CLIPLayer(12, 768) for i in range(12)
|
79 |
+
])
|
80 |
+
|
81 |
+
self.layernorm = nn.LayerNorm(768)
|
82 |
+
|
83 |
+
def forward(self, tokens: torch.LongTensor) -> torch.FloatTensor:
|
84 |
+
tokens = tokens.type(torch.long)
|
85 |
+
|
86 |
+
# (Batch_Size, Seq_Len) -> (Batch_Size, Seq_Len, Dim)
|
87 |
+
state = self.embedding(tokens)
|
88 |
+
|
89 |
+
# Apply encoder layers similar to the Transformer's encoder.
|
90 |
+
for layer in self.layers:
|
91 |
+
# (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
|
92 |
+
state = layer(state)
|
93 |
+
# (Batch_Size, Seq_Len, Dim) -> (Batch_Size, Seq_Len, Dim)
|
94 |
+
output = self.layernorm(state)
|
95 |
+
|
96 |
+
return output
|
SD/ddpm.py
ADDED
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import numpy as np
|
3 |
+
|
4 |
+
class DDPMSampler:
|
5 |
+
|
6 |
+
def __init__(self, generator: torch.Generator, num_training_steps=1000, beta_start: float = 0.00085, beta_end: float = 0.0120):
|
7 |
+
# Params "beta_start" and "beta_end" taken from: https://github.com/CompVis/stable-diffusion/blob/21f890f9da3cfbeaba8e2ac3c425ee9e998d5229/configs/stable-diffusion/v1-inference.yaml#L5C8-L5C8
|
8 |
+
# For the naming conventions, refer to the DDPM paper (https://arxiv.org/pdf/2006.11239.pdf)
|
9 |
+
self.betas = torch.linspace(beta_start ** 0.5, beta_end ** 0.5, num_training_steps, dtype=torch.float32) ** 2
|
10 |
+
self.alphas = 1.0 - self.betas
|
11 |
+
self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
|
12 |
+
self.one = torch.tensor(1.0)
|
13 |
+
|
14 |
+
self.generator = generator
|
15 |
+
|
16 |
+
self.num_train_timesteps = num_training_steps
|
17 |
+
self.timesteps = torch.from_numpy(np.arange(0, num_training_steps)[::-1].copy())
|
18 |
+
|
19 |
+
def set_inference_timesteps(self, num_inference_steps=50):
|
20 |
+
self.num_inference_steps = num_inference_steps
|
21 |
+
step_ratio = self.num_train_timesteps // self.num_inference_steps
|
22 |
+
timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
|
23 |
+
self.timesteps = torch.from_numpy(timesteps)
|
24 |
+
|
25 |
+
def _get_previous_timestep(self, timestep: int) -> int:
|
26 |
+
prev_t = timestep - self.num_train_timesteps // self.num_inference_steps
|
27 |
+
return prev_t
|
28 |
+
|
29 |
+
def _get_variance(self, timestep: int) -> torch.Tensor:
|
30 |
+
prev_t = self._get_previous_timestep(timestep)
|
31 |
+
|
32 |
+
alpha_prod_t = self.alphas_cumprod[timestep]
|
33 |
+
alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one
|
34 |
+
current_beta_t = 1 - alpha_prod_t / alpha_prod_t_prev
|
35 |
+
|
36 |
+
# For t > 0, compute predicted variance βt (see formula (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
|
37 |
+
# and sample from it to get previous sample
|
38 |
+
# x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
|
39 |
+
variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t
|
40 |
+
|
41 |
+
# we always take the log of variance, so clamp it to ensure it's not 0
|
42 |
+
variance = torch.clamp(variance, min=1e-20)
|
43 |
+
|
44 |
+
return variance
|
45 |
+
|
46 |
+
def set_strength(self, strength=1):
|
47 |
+
"""
|
48 |
+
Set how much noise to add to the input image.
|
49 |
+
More noise (strength ~ 1) means that the output will be further from the input image.
|
50 |
+
Less noise (strength ~ 0) means that the output will be closer to the input image.
|
51 |
+
"""
|
52 |
+
# start_step is the number of noise levels to skip
|
53 |
+
start_step = self.num_inference_steps - int(self.num_inference_steps * strength)
|
54 |
+
self.timesteps = self.timesteps[start_step:]
|
55 |
+
self.start_step = start_step
|
56 |
+
|
57 |
+
def step(self, timestep: int, latents: torch.Tensor, model_output: torch.Tensor):
|
58 |
+
t = timestep
|
59 |
+
prev_t = self._get_previous_timestep(t)
|
60 |
+
|
61 |
+
# 1. compute alphas, betas
|
62 |
+
alpha_prod_t = self.alphas_cumprod[t]
|
63 |
+
alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one
|
64 |
+
beta_prod_t = 1 - alpha_prod_t
|
65 |
+
beta_prod_t_prev = 1 - alpha_prod_t_prev
|
66 |
+
current_alpha_t = alpha_prod_t / alpha_prod_t_prev
|
67 |
+
current_beta_t = 1 - current_alpha_t
|
68 |
+
|
69 |
+
# 2. compute predicted original sample from predicted noise also called
|
70 |
+
# "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
|
71 |
+
pred_original_sample = (latents - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
|
72 |
+
|
73 |
+
# 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
|
74 |
+
# See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
|
75 |
+
pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t
|
76 |
+
current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t
|
77 |
+
|
78 |
+
# 5. Compute predicted previous sample µ_t
|
79 |
+
# See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
|
80 |
+
pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * latents
|
81 |
+
|
82 |
+
# 6. Add noise
|
83 |
+
variance = 0
|
84 |
+
if t > 0:
|
85 |
+
device = model_output.device
|
86 |
+
noise = torch.randn(model_output.shape, generator=self.generator, device=device, dtype=model_output.dtype)
|
87 |
+
# Compute the variance as per formula (7) from https://arxiv.org/pdf/2006.11239.pdf
|
88 |
+
variance = (self._get_variance(t) ** 0.5) * noise
|
89 |
+
|
90 |
+
# sample from N(mu, sigma) = X can be obtained by X = mu + sigma * N(0, 1)
|
91 |
+
# the variable "variance" is already multiplied by the noise N(0, 1)
|
92 |
+
pred_prev_sample = pred_prev_sample + variance
|
93 |
+
|
94 |
+
return pred_prev_sample
|
95 |
+
|
96 |
+
def add_noise(
|
97 |
+
self,
|
98 |
+
original_samples: torch.FloatTensor,
|
99 |
+
timesteps: torch.IntTensor,
|
100 |
+
) -> torch.FloatTensor:
|
101 |
+
alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype)
|
102 |
+
timesteps = timesteps.to(original_samples.device)
|
103 |
+
|
104 |
+
sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
|
105 |
+
sqrt_alpha_prod = sqrt_alpha_prod.flatten()
|
106 |
+
while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
|
107 |
+
sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
|
108 |
+
|
109 |
+
sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
|
110 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
|
111 |
+
while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
|
112 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
|
113 |
+
|
114 |
+
# Sample from q(x_t | x_0) as in equation (4) of https://arxiv.org/pdf/2006.11239.pdf
|
115 |
+
# Because N(mu, sigma) = X can be obtained by X = mu + sigma * N(0, 1)
|
116 |
+
# here mu = sqrt_alpha_prod * original_samples and sigma = sqrt_one_minus_alpha_prod
|
117 |
+
noise = torch.randn(original_samples.shape, generator=self.generator, device=original_samples.device, dtype=original_samples.dtype)
|
118 |
+
noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
|
119 |
+
return noisy_samples
|
120 |
+
|
121 |
+
|
122 |
+
|
123 |
+
|
SD/decoder.py
ADDED
@@ -0,0 +1,177 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from attention import SelfAttention
|
5 |
+
|
6 |
+
class VAE_AttentionBlock(nn.Module):
|
7 |
+
def __init__(self, channels):
|
8 |
+
super().__init__()
|
9 |
+
self.groupnorm = nn.GroupNorm(32, channels)
|
10 |
+
self.attention = SelfAttention(1, channels)
|
11 |
+
|
12 |
+
def forward(self, x):
|
13 |
+
# x: (Batch_Size, Features, Height, Width)
|
14 |
+
|
15 |
+
residue = x
|
16 |
+
|
17 |
+
# (Batch_Size, Features, Height, Width) -> (Batch_Size, Features, Height, Width)
|
18 |
+
x = self.groupnorm(x)
|
19 |
+
|
20 |
+
n, c, h, w = x.shape
|
21 |
+
|
22 |
+
# (Batch_Size, Features, Height, Width) -> (Batch_Size, Features, Height * Width)
|
23 |
+
x = x.view((n, c, h * w))
|
24 |
+
|
25 |
+
# (Batch_Size, Features, Height * Width) -> (Batch_Size, Height * Width, Features). Each pixel becomes a feature of size "Features", the sequence length is "Height * Width".
|
26 |
+
x = x.transpose(-1, -2)
|
27 |
+
|
28 |
+
# Perform self-attention WITHOUT mask
|
29 |
+
# (Batch_Size, Height * Width, Features) -> (Batch_Size, Height * Width, Features)
|
30 |
+
x = self.attention(x)
|
31 |
+
|
32 |
+
# (Batch_Size, Height * Width, Features) -> (Batch_Size, Features, Height * Width)
|
33 |
+
x = x.transpose(-1, -2)
|
34 |
+
|
35 |
+
# (Batch_Size, Features, Height * Width) -> (Batch_Size, Features, Height, Width)
|
36 |
+
x = x.view((n, c, h, w))
|
37 |
+
|
38 |
+
# (Batch_Size, Features, Height, Width) + (Batch_Size, Features, Height, Width) -> (Batch_Size, Features, Height, Width)
|
39 |
+
x += residue
|
40 |
+
|
41 |
+
# (Batch_Size, Features, Height, Width)
|
42 |
+
return x
|
43 |
+
|
44 |
+
class VAE_ResidualBlock(nn.Module):
|
45 |
+
def __init__(self, in_channels, out_channels):
|
46 |
+
super().__init__()
|
47 |
+
self.groupnorm_1 = nn.GroupNorm(32, in_channels)
|
48 |
+
self.conv_1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
|
49 |
+
|
50 |
+
self.groupnorm_2 = nn.GroupNorm(32, out_channels)
|
51 |
+
self.conv_2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
|
52 |
+
|
53 |
+
if in_channels == out_channels:
|
54 |
+
self.residual_layer = nn.Identity()
|
55 |
+
else:
|
56 |
+
self.residual_layer = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0)
|
57 |
+
|
58 |
+
def forward(self, x):
|
59 |
+
# x: (Batch_Size, In_Channels, Height, Width)
|
60 |
+
|
61 |
+
residue = x
|
62 |
+
|
63 |
+
# (Batch_Size, In_Channels, Height, Width) -> (Batch_Size, In_Channels, Height, Width)
|
64 |
+
x = self.groupnorm_1(x)
|
65 |
+
|
66 |
+
# (Batch_Size, In_Channels, Height, Width) -> (Batch_Size, In_Channels, Height, Width)
|
67 |
+
x = F.silu(x)
|
68 |
+
|
69 |
+
# (Batch_Size, In_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
70 |
+
x = self.conv_1(x)
|
71 |
+
|
72 |
+
# (Batch_Size, Out_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
73 |
+
x = self.groupnorm_2(x)
|
74 |
+
|
75 |
+
# (Batch_Size, Out_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
76 |
+
x = F.silu(x)
|
77 |
+
|
78 |
+
# (Batch_Size, Out_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
79 |
+
x = self.conv_2(x)
|
80 |
+
|
81 |
+
# (Batch_Size, Out_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
82 |
+
return x + self.residual_layer(residue)
|
83 |
+
|
84 |
+
class VAE_Decoder(nn.Sequential):
|
85 |
+
def __init__(self):
|
86 |
+
super().__init__(
|
87 |
+
# (Batch_Size, 4, Height / 8, Width / 8) -> (Batch_Size, 4, Height / 8, Width / 8)
|
88 |
+
nn.Conv2d(4, 4, kernel_size=1, padding=0),
|
89 |
+
|
90 |
+
# (Batch_Size, 4, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
91 |
+
nn.Conv2d(4, 512, kernel_size=3, padding=1),
|
92 |
+
|
93 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
94 |
+
VAE_ResidualBlock(512, 512),
|
95 |
+
|
96 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
97 |
+
VAE_AttentionBlock(512),
|
98 |
+
|
99 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
100 |
+
VAE_ResidualBlock(512, 512),
|
101 |
+
|
102 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
103 |
+
VAE_ResidualBlock(512, 512),
|
104 |
+
|
105 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
106 |
+
VAE_ResidualBlock(512, 512),
|
107 |
+
|
108 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
109 |
+
VAE_ResidualBlock(512, 512),
|
110 |
+
|
111 |
+
# Repeats the rows and columns of the data by scale_factor (like when you resize an image by doubling its size).
|
112 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 4, Width / 4)
|
113 |
+
nn.Upsample(scale_factor=2),
|
114 |
+
|
115 |
+
# (Batch_Size, 512, Height / 4, Width / 4) -> (Batch_Size, 512, Height / 4, Width / 4)
|
116 |
+
nn.Conv2d(512, 512, kernel_size=3, padding=1),
|
117 |
+
|
118 |
+
# (Batch_Size, 512, Height / 4, Width / 4) -> (Batch_Size, 512, Height / 4, Width / 4)
|
119 |
+
VAE_ResidualBlock(512, 512),
|
120 |
+
|
121 |
+
# (Batch_Size, 512, Height / 4, Width / 4) -> (Batch_Size, 512, Height / 4, Width / 4)
|
122 |
+
VAE_ResidualBlock(512, 512),
|
123 |
+
|
124 |
+
# (Batch_Size, 512, Height / 4, Width / 4) -> (Batch_Size, 512, Height / 4, Width / 4)
|
125 |
+
VAE_ResidualBlock(512, 512),
|
126 |
+
|
127 |
+
# (Batch_Size, 512, Height / 4, Width / 4) -> (Batch_Size, 512, Height / 2, Width / 2)
|
128 |
+
nn.Upsample(scale_factor=2),
|
129 |
+
|
130 |
+
# (Batch_Size, 512, Height / 2, Width / 2) -> (Batch_Size, 512, Height / 2, Width / 2)
|
131 |
+
nn.Conv2d(512, 512, kernel_size=3, padding=1),
|
132 |
+
|
133 |
+
# (Batch_Size, 512, Height / 2, Width / 2) -> (Batch_Size, 256, Height / 2, Width / 2)
|
134 |
+
VAE_ResidualBlock(512, 256),
|
135 |
+
|
136 |
+
# (Batch_Size, 256, Height / 2, Width / 2) -> (Batch_Size, 256, Height / 2, Width / 2)
|
137 |
+
VAE_ResidualBlock(256, 256),
|
138 |
+
|
139 |
+
# (Batch_Size, 256, Height / 2, Width / 2) -> (Batch_Size, 256, Height / 2, Width / 2)
|
140 |
+
VAE_ResidualBlock(256, 256),
|
141 |
+
|
142 |
+
# (Batch_Size, 256, Height / 2, Width / 2) -> (Batch_Size, 256, Height, Width)
|
143 |
+
nn.Upsample(scale_factor=2),
|
144 |
+
|
145 |
+
# (Batch_Size, 256, Height, Width) -> (Batch_Size, 256, Height, Width)
|
146 |
+
nn.Conv2d(256, 256, kernel_size=3, padding=1),
|
147 |
+
|
148 |
+
# (Batch_Size, 256, Height, Width) -> (Batch_Size, 128, Height, Width)
|
149 |
+
VAE_ResidualBlock(256, 128),
|
150 |
+
|
151 |
+
# (Batch_Size, 128, Height, Width) -> (Batch_Size, 128, Height, Width)
|
152 |
+
VAE_ResidualBlock(128, 128),
|
153 |
+
|
154 |
+
# (Batch_Size, 128, Height, Width) -> (Batch_Size, 128, Height, Width)
|
155 |
+
VAE_ResidualBlock(128, 128),
|
156 |
+
|
157 |
+
# (Batch_Size, 128, Height, Width) -> (Batch_Size, 128, Height, Width)
|
158 |
+
nn.GroupNorm(32, 128),
|
159 |
+
|
160 |
+
# (Batch_Size, 128, Height, Width) -> (Batch_Size, 128, Height, Width)
|
161 |
+
nn.SiLU(),
|
162 |
+
|
163 |
+
# (Batch_Size, 128, Height, Width) -> (Batch_Size, 3, Height, Width)
|
164 |
+
nn.Conv2d(128, 3, kernel_size=3, padding=1),
|
165 |
+
)
|
166 |
+
|
167 |
+
def forward(self, x):
|
168 |
+
# x: (Batch_Size, 4, Height / 8, Width / 8)
|
169 |
+
|
170 |
+
# Remove the scaling added by the Encoder.
|
171 |
+
x /= 0.18215
|
172 |
+
|
173 |
+
for module in self:
|
174 |
+
x = module(x)
|
175 |
+
|
176 |
+
# (Batch_Size, 3, Height, Width)
|
177 |
+
return x
|
SD/diffusion.py
ADDED
@@ -0,0 +1,349 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from attention import SelfAttention, CrossAttention
|
5 |
+
|
6 |
+
class TimeEmbedding(nn.Module):
|
7 |
+
def __init__(self, n_embd):
|
8 |
+
super().__init__()
|
9 |
+
self.linear_1 = nn.Linear(n_embd, 4 * n_embd)
|
10 |
+
self.linear_2 = nn.Linear(4 * n_embd, 4 * n_embd)
|
11 |
+
|
12 |
+
def forward(self, x):
|
13 |
+
# x: (1, 320)
|
14 |
+
|
15 |
+
# (1, 320) -> (1, 1280)
|
16 |
+
x = self.linear_1(x)
|
17 |
+
|
18 |
+
# (1, 1280) -> (1, 1280)
|
19 |
+
x = F.silu(x)
|
20 |
+
|
21 |
+
# (1, 1280) -> (1, 1280)
|
22 |
+
x = self.linear_2(x)
|
23 |
+
|
24 |
+
return x
|
25 |
+
|
26 |
+
class UNET_ResidualBlock(nn.Module):
|
27 |
+
def __init__(self, in_channels, out_channels, n_time=1280):
|
28 |
+
super().__init__()
|
29 |
+
self.groupnorm_feature = nn.GroupNorm(32, in_channels)
|
30 |
+
self.conv_feature = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
|
31 |
+
self.linear_time = nn.Linear(n_time, out_channels)
|
32 |
+
|
33 |
+
self.groupnorm_merged = nn.GroupNorm(32, out_channels)
|
34 |
+
self.conv_merged = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
|
35 |
+
|
36 |
+
if in_channels == out_channels:
|
37 |
+
self.residual_layer = nn.Identity()
|
38 |
+
else:
|
39 |
+
self.residual_layer = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0)
|
40 |
+
|
41 |
+
def forward(self, feature, time):
|
42 |
+
# feature: (Batch_Size, In_Channels, Height, Width)
|
43 |
+
# time: (1, 1280)
|
44 |
+
|
45 |
+
residue = feature
|
46 |
+
|
47 |
+
# (Batch_Size, In_Channels, Height, Width) -> (Batch_Size, In_Channels, Height, Width)
|
48 |
+
feature = self.groupnorm_feature(feature)
|
49 |
+
|
50 |
+
# (Batch_Size, In_Channels, Height, Width) -> (Batch_Size, In_Channels, Height, Width)
|
51 |
+
feature = F.silu(feature)
|
52 |
+
|
53 |
+
# (Batch_Size, In_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
54 |
+
feature = self.conv_feature(feature)
|
55 |
+
|
56 |
+
# (1, 1280) -> (1, 1280)
|
57 |
+
time = F.silu(time)
|
58 |
+
|
59 |
+
# (1, 1280) -> (1, Out_Channels)
|
60 |
+
time = self.linear_time(time)
|
61 |
+
|
62 |
+
# Add width and height dimension to time.
|
63 |
+
# (Batch_Size, Out_Channels, Height, Width) + (1, Out_Channels, 1, 1) -> (Batch_Size, Out_Channels, Height, Width)
|
64 |
+
merged = feature + time.unsqueeze(-1).unsqueeze(-1)
|
65 |
+
|
66 |
+
# (Batch_Size, Out_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
67 |
+
merged = self.groupnorm_merged(merged)
|
68 |
+
|
69 |
+
# (Batch_Size, Out_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
70 |
+
merged = F.silu(merged)
|
71 |
+
|
72 |
+
# (Batch_Size, Out_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
73 |
+
merged = self.conv_merged(merged)
|
74 |
+
|
75 |
+
# (Batch_Size, Out_Channels, Height, Width) + (Batch_Size, Out_Channels, Height, Width) -> (Batch_Size, Out_Channels, Height, Width)
|
76 |
+
return merged + self.residual_layer(residue)
|
77 |
+
|
78 |
+
class UNET_AttentionBlock(nn.Module):
|
79 |
+
def __init__(self, n_head: int, n_embd: int, d_context=768):
|
80 |
+
super().__init__()
|
81 |
+
channels = n_head * n_embd
|
82 |
+
|
83 |
+
self.groupnorm = nn.GroupNorm(32, channels, eps=1e-6)
|
84 |
+
self.conv_input = nn.Conv2d(channels, channels, kernel_size=1, padding=0)
|
85 |
+
|
86 |
+
self.layernorm_1 = nn.LayerNorm(channels)
|
87 |
+
self.attention_1 = SelfAttention(n_head, channels, in_proj_bias=False)
|
88 |
+
self.layernorm_2 = nn.LayerNorm(channels)
|
89 |
+
self.attention_2 = CrossAttention(n_head, channels, d_context, in_proj_bias=False)
|
90 |
+
self.layernorm_3 = nn.LayerNorm(channels)
|
91 |
+
self.linear_geglu_1 = nn.Linear(channels, 4 * channels * 2)
|
92 |
+
self.linear_geglu_2 = nn.Linear(4 * channels, channels)
|
93 |
+
|
94 |
+
self.conv_output = nn.Conv2d(channels, channels, kernel_size=1, padding=0)
|
95 |
+
|
96 |
+
def forward(self, x, context):
|
97 |
+
# x: (Batch_Size, Features, Height, Width)
|
98 |
+
# context: (Batch_Size, Seq_Len, Dim)
|
99 |
+
|
100 |
+
residue_long = x
|
101 |
+
|
102 |
+
# (Batch_Size, Features, Height, Width) -> (Batch_Size, Features, Height, Width)
|
103 |
+
x = self.groupnorm(x)
|
104 |
+
|
105 |
+
# (Batch_Size, Features, Height, Width) -> (Batch_Size, Features, Height, Width)
|
106 |
+
x = self.conv_input(x)
|
107 |
+
|
108 |
+
n, c, h, w = x.shape
|
109 |
+
|
110 |
+
# (Batch_Size, Features, Height, Width) -> (Batch_Size, Features, Height * Width)
|
111 |
+
x = x.view((n, c, h * w))
|
112 |
+
|
113 |
+
# (Batch_Size, Features, Height * Width) -> (Batch_Size, Height * Width, Features)
|
114 |
+
x = x.transpose(-1, -2)
|
115 |
+
|
116 |
+
# Normalization + Self-Attention with skip connection
|
117 |
+
|
118 |
+
# (Batch_Size, Height * Width, Features)
|
119 |
+
residue_short = x
|
120 |
+
|
121 |
+
# (Batch_Size, Height * Width, Features) -> (Batch_Size, Height * Width, Features)
|
122 |
+
x = self.layernorm_1(x)
|
123 |
+
|
124 |
+
# (Batch_Size, Height * Width, Features) -> (Batch_Size, Height * Width, Features)
|
125 |
+
x = self.attention_1(x)
|
126 |
+
|
127 |
+
# (Batch_Size, Height * Width, Features) + (Batch_Size, Height * Width, Features) -> (Batch_Size, Height * Width, Features)
|
128 |
+
x += residue_short
|
129 |
+
|
130 |
+
# (Batch_Size, Height * Width, Features)
|
131 |
+
residue_short = x
|
132 |
+
|
133 |
+
# Normalization + Cross-Attention with skip connection
|
134 |
+
|
135 |
+
# (Batch_Size, Height * Width, Features) -> (Batch_Size, Height * Width, Features)
|
136 |
+
x = self.layernorm_2(x)
|
137 |
+
|
138 |
+
# (Batch_Size, Height * Width, Features) -> (Batch_Size, Height * Width, Features)
|
139 |
+
x = self.attention_2(x, context)
|
140 |
+
|
141 |
+
# (Batch_Size, Height * Width, Features) + (Batch_Size, Height * Width, Features) -> (Batch_Size, Height * Width, Features)
|
142 |
+
x += residue_short
|
143 |
+
|
144 |
+
# (Batch_Size, Height * Width, Features)
|
145 |
+
residue_short = x
|
146 |
+
|
147 |
+
# Normalization + FFN with GeGLU and skip connection
|
148 |
+
|
149 |
+
# (Batch_Size, Height * Width, Features) -> (Batch_Size, Height * Width, Features)
|
150 |
+
x = self.layernorm_3(x)
|
151 |
+
|
152 |
+
# GeGLU as implemented in the original code: https://github.com/CompVis/stable-diffusion/blob/21f890f9da3cfbeaba8e2ac3c425ee9e998d5229/ldm/modules/attention.py#L37C10-L37C10
|
153 |
+
# (Batch_Size, Height * Width, Features) -> two tensors of shape (Batch_Size, Height * Width, Features * 4)
|
154 |
+
x, gate = self.linear_geglu_1(x).chunk(2, dim=-1)
|
155 |
+
|
156 |
+
# Element-wise product: (Batch_Size, Height * Width, Features * 4) * (Batch_Size, Height * Width, Features * 4) -> (Batch_Size, Height * Width, Features * 4)
|
157 |
+
x = x * F.gelu(gate)
|
158 |
+
|
159 |
+
# (Batch_Size, Height * Width, Features * 4) -> (Batch_Size, Height * Width, Features)
|
160 |
+
x = self.linear_geglu_2(x)
|
161 |
+
|
162 |
+
# (Batch_Size, Height * Width, Features) + (Batch_Size, Height * Width, Features) -> (Batch_Size, Height * Width, Features)
|
163 |
+
x += residue_short
|
164 |
+
|
165 |
+
# (Batch_Size, Height * Width, Features) -> (Batch_Size, Features, Height * Width)
|
166 |
+
x = x.transpose(-1, -2)
|
167 |
+
|
168 |
+
# (Batch_Size, Features, Height * Width) -> (Batch_Size, Features, Height, Width)
|
169 |
+
x = x.view((n, c, h, w))
|
170 |
+
|
171 |
+
# Final skip connection between initial input and output of the block
|
172 |
+
# (Batch_Size, Features, Height, Width) + (Batch_Size, Features, Height, Width) -> (Batch_Size, Features, Height, Width)
|
173 |
+
return self.conv_output(x) + residue_long
|
174 |
+
|
175 |
+
class Upsample(nn.Module):
|
176 |
+
def __init__(self, channels):
|
177 |
+
super().__init__()
|
178 |
+
self.conv = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
|
179 |
+
|
180 |
+
def forward(self, x):
|
181 |
+
# (Batch_Size, Features, Height, Width) -> (Batch_Size, Features, Height * 2, Width * 2)
|
182 |
+
x = F.interpolate(x, scale_factor=2, mode='nearest')
|
183 |
+
return self.conv(x)
|
184 |
+
|
185 |
+
class SwitchSequential(nn.Sequential):
|
186 |
+
def forward(self, x, context, time):
|
187 |
+
for layer in self:
|
188 |
+
if isinstance(layer, UNET_AttentionBlock):
|
189 |
+
x = layer(x, context)
|
190 |
+
elif isinstance(layer, UNET_ResidualBlock):
|
191 |
+
x = layer(x, time)
|
192 |
+
else:
|
193 |
+
x = layer(x)
|
194 |
+
return x
|
195 |
+
|
196 |
+
class UNET(nn.Module):
|
197 |
+
def __init__(self):
|
198 |
+
super().__init__()
|
199 |
+
self.encoders = nn.ModuleList([
|
200 |
+
# (Batch_Size, 4, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8)
|
201 |
+
SwitchSequential(nn.Conv2d(4, 320, kernel_size=3, padding=1)),
|
202 |
+
|
203 |
+
# (Batch_Size, 320, Height / 8, Width / 8) -> # (Batch_Size, 320, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8)
|
204 |
+
SwitchSequential(UNET_ResidualBlock(320, 320), UNET_AttentionBlock(8, 40)),
|
205 |
+
|
206 |
+
# (Batch_Size, 320, Height / 8, Width / 8) -> # (Batch_Size, 320, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8)
|
207 |
+
SwitchSequential(UNET_ResidualBlock(320, 320), UNET_AttentionBlock(8, 40)),
|
208 |
+
|
209 |
+
# (Batch_Size, 320, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 16, Width / 16)
|
210 |
+
SwitchSequential(nn.Conv2d(320, 320, kernel_size=3, stride=2, padding=1)),
|
211 |
+
|
212 |
+
# (Batch_Size, 320, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16)
|
213 |
+
SwitchSequential(UNET_ResidualBlock(320, 640), UNET_AttentionBlock(8, 80)),
|
214 |
+
|
215 |
+
# (Batch_Size, 640, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16)
|
216 |
+
SwitchSequential(UNET_ResidualBlock(640, 640), UNET_AttentionBlock(8, 80)),
|
217 |
+
|
218 |
+
# (Batch_Size, 640, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 32, Width / 32)
|
219 |
+
SwitchSequential(nn.Conv2d(640, 640, kernel_size=3, stride=2, padding=1)),
|
220 |
+
|
221 |
+
# (Batch_Size, 640, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32)
|
222 |
+
SwitchSequential(UNET_ResidualBlock(640, 1280), UNET_AttentionBlock(8, 160)),
|
223 |
+
|
224 |
+
# (Batch_Size, 1280, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32)
|
225 |
+
SwitchSequential(UNET_ResidualBlock(1280, 1280), UNET_AttentionBlock(8, 160)),
|
226 |
+
|
227 |
+
# (Batch_Size, 1280, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 64, Width / 64)
|
228 |
+
SwitchSequential(nn.Conv2d(1280, 1280, kernel_size=3, stride=2, padding=1)),
|
229 |
+
|
230 |
+
# (Batch_Size, 1280, Height / 64, Width / 64) -> (Batch_Size, 1280, Height / 64, Width / 64)
|
231 |
+
SwitchSequential(UNET_ResidualBlock(1280, 1280)),
|
232 |
+
|
233 |
+
# (Batch_Size, 1280, Height / 64, Width / 64) -> (Batch_Size, 1280, Height / 64, Width / 64)
|
234 |
+
SwitchSequential(UNET_ResidualBlock(1280, 1280)),
|
235 |
+
])
|
236 |
+
|
237 |
+
self.bottleneck = SwitchSequential(
|
238 |
+
# (Batch_Size, 1280, Height / 64, Width / 64) -> (Batch_Size, 1280, Height / 64, Width / 64)
|
239 |
+
UNET_ResidualBlock(1280, 1280),
|
240 |
+
|
241 |
+
# (Batch_Size, 1280, Height / 64, Width / 64) -> (Batch_Size, 1280, Height / 64, Width / 64)
|
242 |
+
UNET_AttentionBlock(8, 160),
|
243 |
+
|
244 |
+
# (Batch_Size, 1280, Height / 64, Width / 64) -> (Batch_Size, 1280, Height / 64, Width / 64)
|
245 |
+
UNET_ResidualBlock(1280, 1280),
|
246 |
+
)
|
247 |
+
|
248 |
+
self.decoders = nn.ModuleList([
|
249 |
+
# (Batch_Size, 2560, Height / 64, Width / 64) -> (Batch_Size, 1280, Height / 64, Width / 64)
|
250 |
+
SwitchSequential(UNET_ResidualBlock(2560, 1280)),
|
251 |
+
|
252 |
+
# (Batch_Size, 2560, Height / 64, Width / 64) -> (Batch_Size, 1280, Height / 64, Width / 64)
|
253 |
+
SwitchSequential(UNET_ResidualBlock(2560, 1280)),
|
254 |
+
|
255 |
+
# (Batch_Size, 2560, Height / 64, Width / 64) -> (Batch_Size, 1280, Height / 64, Width / 64) -> (Batch_Size, 1280, Height / 32, Width / 32)
|
256 |
+
SwitchSequential(UNET_ResidualBlock(2560, 1280), Upsample(1280)),
|
257 |
+
|
258 |
+
# (Batch_Size, 2560, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32)
|
259 |
+
SwitchSequential(UNET_ResidualBlock(2560, 1280), UNET_AttentionBlock(8, 160)),
|
260 |
+
|
261 |
+
# (Batch_Size, 2560, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32)
|
262 |
+
SwitchSequential(UNET_ResidualBlock(2560, 1280), UNET_AttentionBlock(8, 160)),
|
263 |
+
|
264 |
+
# (Batch_Size, 1920, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 32, Width / 32) -> (Batch_Size, 1280, Height / 16, Width / 16)
|
265 |
+
SwitchSequential(UNET_ResidualBlock(1920, 1280), UNET_AttentionBlock(8, 160), Upsample(1280)),
|
266 |
+
|
267 |
+
# (Batch_Size, 1920, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16)
|
268 |
+
SwitchSequential(UNET_ResidualBlock(1920, 640), UNET_AttentionBlock(8, 80)),
|
269 |
+
|
270 |
+
# (Batch_Size, 1280, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16)
|
271 |
+
SwitchSequential(UNET_ResidualBlock(1280, 640), UNET_AttentionBlock(8, 80)),
|
272 |
+
|
273 |
+
# (Batch_Size, 960, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 16, Width / 16) -> (Batch_Size, 640, Height / 8, Width / 8)
|
274 |
+
SwitchSequential(UNET_ResidualBlock(960, 640), UNET_AttentionBlock(8, 80), Upsample(640)),
|
275 |
+
|
276 |
+
# (Batch_Size, 960, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8)
|
277 |
+
SwitchSequential(UNET_ResidualBlock(960, 320), UNET_AttentionBlock(8, 40)),
|
278 |
+
|
279 |
+
# (Batch_Size, 640, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8)
|
280 |
+
SwitchSequential(UNET_ResidualBlock(640, 320), UNET_AttentionBlock(8, 40)),
|
281 |
+
|
282 |
+
# (Batch_Size, 640, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8)
|
283 |
+
SwitchSequential(UNET_ResidualBlock(640, 320), UNET_AttentionBlock(8, 40)),
|
284 |
+
])
|
285 |
+
|
286 |
+
def forward(self, x, context, time):
|
287 |
+
# x: (Batch_Size, 4, Height / 8, Width / 8)
|
288 |
+
# context: (Batch_Size, Seq_Len, Dim)
|
289 |
+
# time: (1, 1280)
|
290 |
+
|
291 |
+
skip_connections = []
|
292 |
+
for layers in self.encoders:
|
293 |
+
x = layers(x, context, time)
|
294 |
+
skip_connections.append(x)
|
295 |
+
|
296 |
+
x = self.bottleneck(x, context, time)
|
297 |
+
|
298 |
+
for layers in self.decoders:
|
299 |
+
# Since we always concat with the skip connection of the encoder, the number of features increases before being sent to the decoder's layer
|
300 |
+
x = torch.cat((x, skip_connections.pop()), dim=1)
|
301 |
+
x = layers(x, context, time)
|
302 |
+
|
303 |
+
return x
|
304 |
+
|
305 |
+
|
306 |
+
class UNET_OutputLayer(nn.Module):
|
307 |
+
def __init__(self, in_channels, out_channels):
|
308 |
+
super().__init__()
|
309 |
+
self.groupnorm = nn.GroupNorm(32, in_channels)
|
310 |
+
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
|
311 |
+
|
312 |
+
def forward(self, x):
|
313 |
+
# x: (Batch_Size, 320, Height / 8, Width / 8)
|
314 |
+
|
315 |
+
# (Batch_Size, 320, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8)
|
316 |
+
x = self.groupnorm(x)
|
317 |
+
|
318 |
+
# (Batch_Size, 320, Height / 8, Width / 8) -> (Batch_Size, 320, Height / 8, Width / 8)
|
319 |
+
x = F.silu(x)
|
320 |
+
|
321 |
+
# (Batch_Size, 320, Height / 8, Width / 8) -> (Batch_Size, 4, Height / 8, Width / 8)
|
322 |
+
x = self.conv(x)
|
323 |
+
|
324 |
+
# (Batch_Size, 4, Height / 8, Width / 8)
|
325 |
+
return x
|
326 |
+
|
327 |
+
class Diffusion(nn.Module):
|
328 |
+
def __init__(self):
|
329 |
+
super().__init__()
|
330 |
+
self.time_embedding = TimeEmbedding(320)
|
331 |
+
self.unet = UNET()
|
332 |
+
self.final = UNET_OutputLayer(320, 4)
|
333 |
+
|
334 |
+
def forward(self, latent, context, time):
|
335 |
+
# latent: (Batch_Size, 4, Height / 8, Width / 8)
|
336 |
+
# context: (Batch_Size, Seq_Len, Dim)
|
337 |
+
# time: (1, 320)
|
338 |
+
|
339 |
+
# (1, 320) -> (1, 1280)
|
340 |
+
time = self.time_embedding(time)
|
341 |
+
|
342 |
+
# (Batch, 4, Height / 8, Width / 8) -> (Batch, 320, Height / 8, Width / 8)
|
343 |
+
output = self.unet(latent, context, time)
|
344 |
+
|
345 |
+
# (Batch, 320, Height / 8, Width / 8) -> (Batch, 4, Height / 8, Width / 8)
|
346 |
+
output = self.final(output)
|
347 |
+
|
348 |
+
# (Batch, 4, Height / 8, Width / 8)
|
349 |
+
return output
|
SD/encoder.py
ADDED
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from decoder import VAE_AttentionBlock, VAE_ResidualBlock
|
5 |
+
|
6 |
+
class VAE_Encoder(nn.Sequential):
|
7 |
+
def __init__(self):
|
8 |
+
super().__init__(
|
9 |
+
# (Batch_Size, Channel, Height, Width) -> (Batch_Size, 128, Height, Width)
|
10 |
+
nn.Conv2d(3, 128, kernel_size=3, padding=1),
|
11 |
+
|
12 |
+
# (Batch_Size, 128, Height, Width) -> (Batch_Size, 128, Height, Width)
|
13 |
+
VAE_ResidualBlock(128, 128),
|
14 |
+
|
15 |
+
# (Batch_Size, 128, Height, Width) -> (Batch_Size, 128, Height, Width)
|
16 |
+
VAE_ResidualBlock(128, 128),
|
17 |
+
|
18 |
+
# (Batch_Size, 128, Height, Width) -> (Batch_Size, 128, Height / 2, Width / 2)
|
19 |
+
nn.Conv2d(128, 128, kernel_size=3, stride=2, padding=0),
|
20 |
+
|
21 |
+
# (Batch_Size, 128, Height / 2, Width / 2) -> (Batch_Size, 256, Height / 2, Width / 2)
|
22 |
+
VAE_ResidualBlock(128, 256),
|
23 |
+
|
24 |
+
# (Batch_Size, 256, Height / 2, Width / 2) -> (Batch_Size, 256, Height / 2, Width / 2)
|
25 |
+
VAE_ResidualBlock(256, 256),
|
26 |
+
|
27 |
+
# (Batch_Size, 256, Height / 2, Width / 2) -> (Batch_Size, 256, Height / 4, Width / 4)
|
28 |
+
nn.Conv2d(256, 256, kernel_size=3, stride=2, padding=0),
|
29 |
+
|
30 |
+
# (Batch_Size, 256, Height / 4, Width / 4) -> (Batch_Size, 512, Height / 4, Width / 4)
|
31 |
+
VAE_ResidualBlock(256, 512),
|
32 |
+
|
33 |
+
# (Batch_Size, 512, Height / 4, Width / 4) -> (Batch_Size, 512, Height / 4, Width / 4)
|
34 |
+
VAE_ResidualBlock(512, 512),
|
35 |
+
|
36 |
+
# (Batch_Size, 512, Height / 4, Width / 4) -> (Batch_Size, 512, Height / 8, Width / 8)
|
37 |
+
nn.Conv2d(512, 512, kernel_size=3, stride=2, padding=0),
|
38 |
+
|
39 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
40 |
+
VAE_ResidualBlock(512, 512),
|
41 |
+
|
42 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
43 |
+
VAE_ResidualBlock(512, 512),
|
44 |
+
|
45 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
46 |
+
VAE_ResidualBlock(512, 512),
|
47 |
+
|
48 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
49 |
+
VAE_AttentionBlock(512),
|
50 |
+
|
51 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
52 |
+
VAE_ResidualBlock(512, 512),
|
53 |
+
|
54 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
55 |
+
nn.GroupNorm(32, 512),
|
56 |
+
|
57 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 512, Height / 8, Width / 8)
|
58 |
+
nn.SiLU(),
|
59 |
+
|
60 |
+
# Because the padding=1, it means the width and height will increase by 2
|
61 |
+
# Out_Height = In_Height + Padding_Top + Padding_Bottom
|
62 |
+
# Out_Width = In_Width + Padding_Left + Padding_Right
|
63 |
+
# Since padding = 1 means Padding_Top = Padding_Bottom = Padding_Left = Padding_Right = 1,
|
64 |
+
# Since the Out_Width = In_Width + 2 (same for Out_Height), it will compensate for the Kernel size of 3
|
65 |
+
# (Batch_Size, 512, Height / 8, Width / 8) -> (Batch_Size, 8, Height / 8, Width / 8).
|
66 |
+
nn.Conv2d(512, 8, kernel_size=3, padding=1),
|
67 |
+
|
68 |
+
# (Batch_Size, 8, Height / 8, Width / 8) -> (Batch_Size, 8, Height / 8, Width / 8)
|
69 |
+
nn.Conv2d(8, 8, kernel_size=1, padding=0),
|
70 |
+
)
|
71 |
+
|
72 |
+
def forward(self, x, noise):
|
73 |
+
# x: (Batch_Size, Channel, Height, Width)
|
74 |
+
# noise: (Batch_Size, 4, Height / 8, Width / 8)
|
75 |
+
|
76 |
+
for module in self:
|
77 |
+
|
78 |
+
if getattr(module, 'stride', None) == (2, 2): # Padding at downsampling should be asymmetric (see #8)
|
79 |
+
# Pad: (Padding_Left, Padding_Right, Padding_Top, Padding_Bottom).
|
80 |
+
# Pad with zeros on the right and bottom.
|
81 |
+
# (Batch_Size, Channel, Height, Width) -> (Batch_Size, Channel, Height + Padding_Top + Padding_Bottom, Width + Padding_Left + Padding_Right) = (Batch_Size, Channel, Height + 1, Width + 1)
|
82 |
+
x = F.pad(x, (0, 1, 0, 1))
|
83 |
+
|
84 |
+
x = module(x)
|
85 |
+
# (Batch_Size, 8, Height / 8, Width / 8) -> two tensors of shape (Batch_Size, 4, Height / 8, Width / 8)
|
86 |
+
mean, log_variance = torch.chunk(x, 2, dim=1)
|
87 |
+
# Clamp the log variance between -30 and 20, so that the variance is between (circa) 1e-14 and 1e8.
|
88 |
+
# (Batch_Size, 4, Height / 8, Width / 8) -> (Batch_Size, 4, Height / 8, Width / 8)
|
89 |
+
log_variance = torch.clamp(log_variance, -30, 20)
|
90 |
+
# (Batch_Size, 4, Height / 8, Width / 8) -> (Batch_Size, 4, Height / 8, Width / 8)
|
91 |
+
variance = log_variance.exp()
|
92 |
+
# (Batch_Size, 4, Height / 8, Width / 8) -> (Batch_Size, 4, Height / 8, Width / 8)
|
93 |
+
stdev = variance.sqrt()
|
94 |
+
|
95 |
+
# Transform N(0, 1) -> N(mean, stdev)
|
96 |
+
# (Batch_Size, 4, Height / 8, Width / 8) -> (Batch_Size, 4, Height / 8, Width / 8)
|
97 |
+
x = mean + stdev * noise
|
98 |
+
|
99 |
+
# Scale by a constant
|
100 |
+
# Constant taken from: https://github.com/CompVis/stable-diffusion/blob/21f890f9da3cfbeaba8e2ac3c425ee9e998d5229/configs/stable-diffusion/v1-inference.yaml#L17C1-L17C1
|
101 |
+
x *= 0.18215
|
102 |
+
|
103 |
+
return x
|
SD/model_converter.py
ADDED
The diff for this file is too large to render.
See raw diff
|
|
SD/model_loader.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from clip import CLIP
|
2 |
+
from encoder import VAE_Encoder
|
3 |
+
from decoder import VAE_Decoder
|
4 |
+
from diffusion import Diffusion
|
5 |
+
|
6 |
+
import model_converter
|
7 |
+
|
8 |
+
def preload_models_from_standard_weights(ckpt_path, device):
|
9 |
+
state_dict = model_converter.load_from_standard_weights(ckpt_path, device)
|
10 |
+
|
11 |
+
encoder = VAE_Encoder().to(device)
|
12 |
+
encoder.load_state_dict(state_dict['encoder'], strict=True)
|
13 |
+
|
14 |
+
decoder = VAE_Decoder().to(device)
|
15 |
+
decoder.load_state_dict(state_dict['decoder'], strict=True)
|
16 |
+
|
17 |
+
diffusion = Diffusion().to(device)
|
18 |
+
diffusion.load_state_dict(state_dict['diffusion'], strict=True)
|
19 |
+
|
20 |
+
clip = CLIP().to(device)
|
21 |
+
clip.load_state_dict(state_dict['clip'], strict=True)
|
22 |
+
|
23 |
+
return {
|
24 |
+
'clip': clip,
|
25 |
+
'encoder': encoder,
|
26 |
+
'decoder': decoder,
|
27 |
+
'diffusion': diffusion,
|
28 |
+
}
|
SD/pipeline.py
ADDED
@@ -0,0 +1,170 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import numpy as np
|
3 |
+
from tqdm import tqdm
|
4 |
+
from ddpm import DDPMSampler
|
5 |
+
|
6 |
+
WIDTH = 512
|
7 |
+
HEIGHT = 512
|
8 |
+
LATENTS_WIDTH = WIDTH // 8
|
9 |
+
LATENTS_HEIGHT = HEIGHT // 8
|
10 |
+
|
11 |
+
def generate(
|
12 |
+
prompt,
|
13 |
+
uncond_prompt=None,
|
14 |
+
input_image=None,
|
15 |
+
strength=0.8,
|
16 |
+
do_cfg=True,
|
17 |
+
cfg_scale=7.5,
|
18 |
+
sampler_name="ddpm",
|
19 |
+
n_inference_steps=50,
|
20 |
+
models={},
|
21 |
+
seed=None,
|
22 |
+
device=None,
|
23 |
+
idle_device=None,
|
24 |
+
tokenizer=None,
|
25 |
+
):
|
26 |
+
with torch.no_grad():
|
27 |
+
if not 0 < strength <= 1:
|
28 |
+
raise ValueError("strength must be between 0 and 1")
|
29 |
+
|
30 |
+
if idle_device:
|
31 |
+
to_idle = lambda x: x.to(idle_device)
|
32 |
+
else:
|
33 |
+
to_idle = lambda x: x
|
34 |
+
|
35 |
+
# Initialize random number generator according to the seed specified
|
36 |
+
generator = torch.Generator(device=device)
|
37 |
+
if seed is None:
|
38 |
+
generator.seed()
|
39 |
+
else:
|
40 |
+
generator.manual_seed(seed)
|
41 |
+
|
42 |
+
clip = models["clip"]
|
43 |
+
clip.to(device)
|
44 |
+
|
45 |
+
if do_cfg:
|
46 |
+
# Convert into a list of length Seq_Len=77
|
47 |
+
cond_tokens = tokenizer.batch_encode_plus(
|
48 |
+
[prompt], padding="max_length", max_length=77
|
49 |
+
).input_ids
|
50 |
+
# (Batch_Size, Seq_Len)
|
51 |
+
cond_tokens = torch.tensor(cond_tokens, dtype=torch.long, device=device)
|
52 |
+
# (Batch_Size, Seq_Len) -> (Batch_Size, Seq_Len, Dim)
|
53 |
+
cond_context = clip(cond_tokens)
|
54 |
+
# Convert into a list of length Seq_Len=77
|
55 |
+
uncond_tokens = tokenizer.batch_encode_plus(
|
56 |
+
[uncond_prompt], padding="max_length", max_length=77
|
57 |
+
).input_ids
|
58 |
+
# (Batch_Size, Seq_Len)
|
59 |
+
uncond_tokens = torch.tensor(uncond_tokens, dtype=torch.long, device=device)
|
60 |
+
# (Batch_Size, Seq_Len) -> (Batch_Size, Seq_Len, Dim)
|
61 |
+
uncond_context = clip(uncond_tokens)
|
62 |
+
# (Batch_Size, Seq_Len, Dim) + (Batch_Size, Seq_Len, Dim) -> (2 * Batch_Size, Seq_Len, Dim)
|
63 |
+
context = torch.cat([cond_context, uncond_context])
|
64 |
+
else:
|
65 |
+
# Convert into a list of length Seq_Len=77
|
66 |
+
tokens = tokenizer.batch_encode_plus(
|
67 |
+
[prompt], padding="max_length", max_length=77
|
68 |
+
).input_ids
|
69 |
+
# (Batch_Size, Seq_Len)
|
70 |
+
tokens = torch.tensor(tokens, dtype=torch.long, device=device)
|
71 |
+
# (Batch_Size, Seq_Len) -> (Batch_Size, Seq_Len, Dim)
|
72 |
+
context = clip(tokens)
|
73 |
+
to_idle(clip)
|
74 |
+
|
75 |
+
if sampler_name == "ddpm":
|
76 |
+
sampler = DDPMSampler(generator)
|
77 |
+
sampler.set_inference_timesteps(n_inference_steps)
|
78 |
+
else:
|
79 |
+
raise ValueError("Unknown sampler value %s. ")
|
80 |
+
|
81 |
+
latents_shape = (1, 4, LATENTS_HEIGHT, LATENTS_WIDTH)
|
82 |
+
|
83 |
+
if input_image:
|
84 |
+
encoder = models["encoder"]
|
85 |
+
encoder.to(device)
|
86 |
+
|
87 |
+
input_image_tensor = input_image.resize((WIDTH, HEIGHT))
|
88 |
+
# (Height, Width, Channel)
|
89 |
+
input_image_tensor = np.array(input_image_tensor)
|
90 |
+
# (Height, Width, Channel) -> (Height, Width, Channel)
|
91 |
+
input_image_tensor = torch.tensor(input_image_tensor, dtype=torch.float32, device=device)
|
92 |
+
# (Height, Width, Channel) -> (Height, Width, Channel)
|
93 |
+
input_image_tensor = rescale(input_image_tensor, (0, 255), (-1, 1))
|
94 |
+
# (Height, Width, Channel) -> (Batch_Size, Height, Width, Channel)
|
95 |
+
input_image_tensor = input_image_tensor.unsqueeze(0)
|
96 |
+
# (Batch_Size, Height, Width, Channel) -> (Batch_Size, Channel, Height, Width)
|
97 |
+
input_image_tensor = input_image_tensor.permute(0, 3, 1, 2)
|
98 |
+
|
99 |
+
# (Batch_Size, 4, Latents_Height, Latents_Width)
|
100 |
+
encoder_noise = torch.randn(latents_shape, generator=generator, device=device)
|
101 |
+
# (Batch_Size, 4, Latents_Height, Latents_Width)
|
102 |
+
latents = encoder(input_image_tensor, encoder_noise)
|
103 |
+
|
104 |
+
# Add noise to the latents (the encoded input image)
|
105 |
+
# (Batch_Size, 4, Latents_Height, Latents_Width)
|
106 |
+
sampler.set_strength(strength=strength)
|
107 |
+
latents = sampler.add_noise(latents, sampler.timesteps[0])
|
108 |
+
|
109 |
+
to_idle(encoder)
|
110 |
+
else:
|
111 |
+
# (Batch_Size, 4, Latents_Height, Latents_Width)
|
112 |
+
latents = torch.randn(latents_shape, generator=generator, device=device)
|
113 |
+
|
114 |
+
diffusion = models["diffusion"]
|
115 |
+
diffusion.to(device)
|
116 |
+
|
117 |
+
timesteps = tqdm(sampler.timesteps)
|
118 |
+
for i, timestep in enumerate(timesteps):
|
119 |
+
# (1, 320)
|
120 |
+
time_embedding = get_time_embedding(timestep).to(device)
|
121 |
+
|
122 |
+
# (Batch_Size, 4, Latents_Height, Latents_Width)
|
123 |
+
model_input = latents
|
124 |
+
|
125 |
+
if do_cfg:
|
126 |
+
# (Batch_Size, 4, Latents_Height, Latents_Width) -> (2 * Batch_Size, 4, Latents_Height, Latents_Width)
|
127 |
+
model_input = model_input.repeat(2, 1, 1, 1)
|
128 |
+
|
129 |
+
# model_output is the predicted noise
|
130 |
+
# (Batch_Size, 4, Latents_Height, Latents_Width) -> (Batch_Size, 4, Latents_Height, Latents_Width)
|
131 |
+
model_output = diffusion(model_input, context, time_embedding)
|
132 |
+
|
133 |
+
if do_cfg:
|
134 |
+
output_cond, output_uncond = model_output.chunk(2)
|
135 |
+
model_output = cfg_scale * (output_cond - output_uncond) + output_uncond
|
136 |
+
|
137 |
+
# (Batch_Size, 4, Latents_Height, Latents_Width) -> (Batch_Size, 4, Latents_Height, Latents_Width)
|
138 |
+
latents = sampler.step(timestep, latents, model_output)
|
139 |
+
|
140 |
+
to_idle(diffusion)
|
141 |
+
|
142 |
+
decoder = models["decoder"]
|
143 |
+
decoder.to(device)
|
144 |
+
# (Batch_Size, 4, Latents_Height, Latents_Width) -> (Batch_Size, 3, Height, Width)
|
145 |
+
images = decoder(latents)
|
146 |
+
to_idle(decoder)
|
147 |
+
|
148 |
+
images = rescale(images, (-1, 1), (0, 255), clamp=True)
|
149 |
+
# (Batch_Size, Channel, Height, Width) -> (Batch_Size, Height, Width, Channel)
|
150 |
+
images = images.permute(0, 2, 3, 1)
|
151 |
+
images = images.to("cpu", torch.uint8).numpy()
|
152 |
+
return images[0]
|
153 |
+
|
154 |
+
def rescale(x, old_range, new_range, clamp=False):
|
155 |
+
old_min, old_max = old_range
|
156 |
+
new_min, new_max = new_range
|
157 |
+
x -= old_min
|
158 |
+
x *= (new_max - new_min) / (old_max - old_min)
|
159 |
+
x += new_min
|
160 |
+
if clamp:
|
161 |
+
x = x.clamp(new_min, new_max)
|
162 |
+
return x
|
163 |
+
|
164 |
+
def get_time_embedding(timestep):
|
165 |
+
# Shape: (160,)
|
166 |
+
freqs = torch.pow(10000, -torch.arange(start=0, end=160, dtype=torch.float32) / 160)
|
167 |
+
# Shape: (1, 160)
|
168 |
+
x = torch.tensor([timestep], dtype=torch.float32)[:, None] * freqs[None]
|
169 |
+
# Shape: (1, 160 * 2)
|
170 |
+
return torch.cat([torch.cos(x), torch.sin(x)], dim=-1)
|
SD/run.py
ADDED
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import model_loader
|
2 |
+
import pipeline
|
3 |
+
from PIL import Image
|
4 |
+
from pathlib import Path
|
5 |
+
from transformers import CLIPTokenizer
|
6 |
+
import torch
|
7 |
+
|
8 |
+
|
9 |
+
DEVICE = "cpu"
|
10 |
+
|
11 |
+
ALLOW_CUDA = True
|
12 |
+
ALLOW_MPS = False
|
13 |
+
|
14 |
+
if torch.cuda.is_available() and ALLOW_CUDA:
|
15 |
+
DEVICE = "cuda"
|
16 |
+
|
17 |
+
print(f"Using device: {DEVICE}")
|
18 |
+
|
19 |
+
tokenizer = CLIPTokenizer("../data/tokenizer_vocab.json", merges_file="../data/tokenizer_merges.txt")
|
20 |
+
model_file = "../data/v1-5-pruned-emaonly.ckpt"
|
21 |
+
models = model_loader.preload_models_from_standard_weights(model_file, device=DEVICE)
|
22 |
+
|
23 |
+
## TEXT TO IMAGE
|
24 |
+
|
25 |
+
# prompt = "A dog with sunglasses, wearing comfy hat, looking at camera, highly detailed, ultra sharp, cinematic, 100mm lens, 8k resolution."
|
26 |
+
prompt = "A cat stretching on the floor, highly detailed, ultra sharp, cinematic, 100mm lens, 8k resolution."
|
27 |
+
uncond_prompt = "" # Also known as negative prom pt
|
28 |
+
do_cfg = True
|
29 |
+
cfg_scale = 8 # min: 1, max: 14
|
30 |
+
|
31 |
+
## IMAGE TO IMAGE
|
32 |
+
|
33 |
+
input_image = None
|
34 |
+
# Comment to disable image to image
|
35 |
+
image_path = "../images/dog.jpg"
|
36 |
+
# input_image = Image.open(image_path)
|
37 |
+
# Higher values means more noise will be added to the input image, so the result will further from the input image.
|
38 |
+
# Lower values means less noise is added to the input image, so output will be closer to the input image.
|
39 |
+
strength = 0.9
|
40 |
+
|
41 |
+
## SAMPLER
|
42 |
+
|
43 |
+
sampler = "ddpm"
|
44 |
+
num_inference_steps = 2
|
45 |
+
seed = 42
|
46 |
+
|
47 |
+
output_image = pipeline.generate(
|
48 |
+
prompt=prompt,
|
49 |
+
uncond_prompt=uncond_prompt,
|
50 |
+
input_image=input_image,
|
51 |
+
strength=strength,
|
52 |
+
do_cfg=do_cfg,
|
53 |
+
cfg_scale=cfg_scale,
|
54 |
+
sampler_name=sampler,
|
55 |
+
n_inference_steps=num_inference_steps,
|
56 |
+
seed=seed,
|
57 |
+
models=models,
|
58 |
+
device=DEVICE,
|
59 |
+
idle_device="cpu",
|
60 |
+
tokenizer=tokenizer,
|
61 |
+
)
|
62 |
+
|
63 |
+
# Combine the input image and the output image into a single image.
|
64 |
+
Image.fromarray(output_image)
|
SD/sd_demo.ipynb
ADDED
The diff for this file is too large to render.
See raw diff
|
|