Spaces:
Build error
Build error
File size: 5,514 Bytes
7931c5f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 |
# ------------------------------------------------------------------------------------
# Minimal DALL-E
# Copyright (c) 2021 KakaoBrain. All Rights Reserved.
# Licensed under the Apache License, Version 2.0 [see LICENSE for details]
# ------------------------------------------------------------------------------------
import torch
from typing import Optional
from tqdm import tqdm
from torch.nn import functional as F
def cutoff_topk_logits(logits: torch.FloatTensor, k: int) -> torch.FloatTensor:
if k is None:
return logits
else:
v, ix = torch.topk(logits, k)
out = logits.clone()
out[out < v[:, [-1]]] = -float('Inf')
return out
def cutoff_topp_probs(probs: torch.FloatTensor, p: float) -> torch.FloatTensor:
if p is None:
return probs
else:
sorted_probs, sorted_indices = torch.sort(probs, dim=-1, descending=True)
cum_probs = torch.cumsum(sorted_probs, dim=-1)
sorted_idx_remove_cond = cum_probs >= p
sorted_idx_remove_cond[..., 1:] = sorted_idx_remove_cond[..., :-1].clone()
sorted_idx_remove_cond[..., 0] = 0
indices_to_remove = sorted_idx_remove_cond.scatter(-1, sorted_indices, sorted_idx_remove_cond)
probs = probs.masked_fill(indices_to_remove, 0.0)
norm_probs = probs / torch.sum(probs, dim=-1, keepdim=True)
return norm_probs
def get_positional_encoding(inputs: torch.LongTensor, mode: str = '1d') -> torch.LongTensor:
device = inputs.device
if mode == '1d':
B, N = inputs.shape
xs_pos = torch.arange(N, device=device).repeat((B, 1))
elif mode == '2d':
B, H, W = inputs.shape
xs_pos_h = torch.arange(H, device=device).repeat(B, W, 1).transpose(1, 2)
xs_pos_w = torch.arange(W, device=device).repeat(B, H, 1)
xs_pos = (xs_pos_h, xs_pos_w)
else:
raise ValueError('%s positional encoding invalid' % mode)
return xs_pos
@torch.no_grad()
def sampling(model: torch.nn.Module,
tokens: torch.LongTensor,
top_k: Optional[float] = None,
top_p: Optional[float] = None,
softmax_temperature: float = 1.0,
is_tqdm: bool = True,
use_fp16: bool = True,
max_seq_len: int = 256) -> torch.LongTensor:
code = None
past = None
pbar = tqdm(range(max_seq_len), total=max_seq_len) if is_tqdm else range(max_seq_len)
pos_enc_tokens = get_positional_encoding(tokens, mode='1d')
for cnt, h in enumerate(pbar):
if code is None:
code_ = None
pos_enc_code_ = None
else:
code_ = code.clone().detach()
pos_enc_code_ = get_positional_encoding(code_, mode='1d')
code_ = code_[:, cnt-1].unsqueeze(-1)
pos_enc_code_ = pos_enc_code_[:, cnt-1].unsqueeze(-1)
logits, present = model.sampling(images=code_,
texts=tokens,
pos_images=pos_enc_code_,
pos_texts=pos_enc_tokens,
use_fp16=use_fp16,
past=past)
logits = logits.to(dtype=torch.float32)
logits = logits / softmax_temperature
present = torch.stack(present).clone().detach()
if past is None:
past = [present]
else:
past.append(present)
logits = cutoff_topk_logits(logits, top_k)
probs = F.softmax(logits, dim=-1)
probs = cutoff_topp_probs(probs, top_p)
idx = torch.multinomial(probs, num_samples=1).clone().detach()
code = idx if code is None else torch.cat([code, idx], axis=1)
del past
return code
@torch.no_grad()
def sampling_igpt(model: torch.nn.Module,
sos: torch.FloatTensor,
top_k: Optional[float] = None,
top_p: Optional[float] = None,
softmax_temperature: float = 1.0,
is_tqdm: bool = True,
use_fp16: bool = True,
max_seq_len: int = 256) -> torch.LongTensor:
code = None
past = None
pbar = tqdm(range(max_seq_len), total=max_seq_len) if is_tqdm else range(max_seq_len)
for cnt, h in enumerate(pbar):
if code is None:
code_ = None
pos_enc_code_ = None
else:
code_ = code.clone().detach()
pos_enc_code_ = get_positional_encoding(code_, mode='1d')
code_ = code_[:, cnt-1].unsqueeze(-1)
pos_enc_code_ = pos_enc_code_[:, cnt-1].unsqueeze(-1)
logits, present = model.sampling(sos=sos,
codes=code_,
pos_codes=pos_enc_code_,
use_fp16=use_fp16,
past=past)
logits = logits.to(dtype=torch.float32)
logits = logits / softmax_temperature
present = torch.stack(present).clone().detach()
if past is None:
past = [present]
else:
past.append(present)
logits = cutoff_topk_logits(logits, top_k)
probs = F.softmax(logits, dim=-1)
probs = cutoff_topp_probs(probs, top_p)
idx = torch.multinomial(probs, num_samples=1).clone().detach()
code = idx if code is None else torch.cat([code, idx], axis=1)
del past
return code
|