Spaces:
Runtime error
Runtime error
File size: 16,478 Bytes
85a5010 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 |
import numpy as np
import torch
import torch.nn.functional as F
import random
from utils import get_init_text, update_token_mask
import time
def generate_step(out, gen_idx, temperature=None, top_k=0, sample=False, return_list=True):
""" Generate a word from out[gen_idx]
args:
- out (torch.Tensor): tensor of logits of size batch_size x seq_len x vocab_size
- gen_idx (int): location for which to generate for
- top_k (int): if >0, only sample from the top k most probable words
- sample (Bool): if True, sample from full distribution. Overridden by top_k
"""
logits = out[:, gen_idx]
if temperature is not None:
logits = logits / temperature
if top_k > 0:
kth_vals, kth_idx = logits.topk(top_k, dim=-1)
dist = torch.distributions.categorical.Categorical(logits=kth_vals)
idx = kth_idx.gather(dim=1, index=dist.sample().unsqueeze(-1)).squeeze(-1)
elif sample:
dist = torch.distributions.categorical.Categorical(logits=logits)
idx = dist.sample().squeeze(-1)
else:
idx = torch.argmax(logits, dim=-1)
return idx.tolist() if return_list else idx
def generate_caption_step(out, gen_idx, mask, temperature=None, top_k=100):
""" Generate a word from out[gen_idx]
args:
- out (torch.Tensor): tensor of logits of size (batch_size, seq_len, vocab_size)
- gen_idx (int): location for which to generate for
- mask (torch.Tensor): (1, vocab_size)
- top_k (int): candidate k
"""
logits = out[:, gen_idx]
if temperature is not None:
logits = logits / temperature
probs = F.softmax(logits, dim=-1)
probs *= (mask)
top_k_probs, top_k_ids = probs.topk(top_k, dim=-1)
return top_k_probs, top_k_ids
def sequential_generation(model, clip, tokenizer, image_instance,token_mask, prompt, logger,
max_len=15, top_k=100,temperature=None, alpha=0.7,beta=1,
max_iters=20,batch_size=1, verbose=True):
""" Generate one word at a time, in L->R order """
seed_len = len(prompt.split())+1
batch = get_init_text(tokenizer, prompt, max_len, batch_size)
image_embeds = clip.compute_image_representation_from_image_instance(image_instance)
clip_score_sequence = []
best_clip_score = 0
inp = torch.tensor(batch).to(image_embeds.device)
gen_texts = []
for iter_num in range(max_iters):
for ii in range(max_len):
token_mask = update_token_mask(tokenizer, token_mask, max_len, ii)
for jj in range(batch_size):
inp[jj][seed_len + ii] = tokenizer.mask_token_id
inp_ = inp.clone().detach()
out = model(inp).logits
probs, idxs = generate_caption_step(out, gen_idx=seed_len + ii,mask=token_mask, top_k=top_k, temperature=temperature)
for jj in range(batch_size):
topk_inp = inp_.repeat(top_k, 1)
idxs_ = (idxs[jj] * token_mask[0][idxs[jj]]).long()
topk_inp[:, ii + seed_len] = idxs_
batch_text_list = tokenizer.batch_decode(topk_inp, skip_special_tokens=True)
clip_score, clip_ref = clip.compute_image_text_similarity_via_raw_text(image_embeds, batch_text_list)
final_score = alpha * probs + beta * clip_score
best_clip_id = final_score.argmax()
inp[jj][seed_len + ii] = idxs_[best_clip_id]
current_clip_score = clip_ref[jj][best_clip_id]
clip_score_sequence.append(current_clip_score.cpu().item())
if verbose and np.mod(iter_num + 1, 1) == 0:
for_print = tokenizer.decode(inp[0])
cur_text = tokenizer.decode(inp[0],skip_special_tokens=True)
if best_clip_score < current_clip_score.cpu().item():
best_clip_score = current_clip_score.cpu().item()
best_caption = cur_text
gen_texts.append(cur_text)
logger.info(f"iter {iter_num + 1}, clip score {current_clip_score:.3f}: "+ for_print)
gen_texts.append(best_caption)
clip_score_sequence.append(best_clip_score)
return gen_texts, clip_score_sequence
def shuffle_generation(model, clip, tokenizer,image_instance,token_mask, prompt, logger,
max_len=15, top_k=0,temperature=None, alpha=0.7,beta=1,
max_iters=20,batch_size=1,
verbose=True):
""" Generate one word at a time, in random generation order """
seed_len = len(prompt.split())+1
batch = get_init_text(tokenizer,prompt, max_len, batch_size)
image_embeds = clip.compute_image_representation_from_image_instance(image_instance)
inp = torch.tensor(batch).to(image_embeds.device)
clip_score_sequence = []
best_clip_score = 0
random_lst = list(range(max_len))
random.shuffle(random_lst)
logger.info(f"Order_list:{random_lst}")
gen_texts = []
for iter_num in range(max_iters):
for ii in random_lst:
token_mask = update_token_mask(tokenizer, token_mask, max_len, ii)
for jj in range(batch_size):
inp[jj][seed_len + ii] = tokenizer.mask_token_id
inp_ = inp.clone().detach()
out = model(inp).logits
probs, idxs = generate_caption_step(out, gen_idx=seed_len + ii,mask=token_mask, top_k=top_k, temperature=temperature)
for jj in range(batch_size):
topk_inp = inp_.repeat(top_k, 1)
topk_inp[:, ii + seed_len] = (idxs[jj] * token_mask[0][idxs[jj]]).long()
batch_text_list = tokenizer.batch_decode(topk_inp, skip_special_tokens=True)
clip_score,clip_ref = clip.compute_image_text_similarity_via_raw_text(image_embeds, batch_text_list)
final_score = alpha * probs + beta * clip_score
best_clip_id = final_score.argmax()
inp[jj][seed_len + ii] = idxs[jj][best_clip_id]
current_clip_score = clip_ref[jj][best_clip_id]
clip_score_sequence.append(current_clip_score.cpu().item())
if verbose and np.mod(iter_num + 1, 1) == 0:
for_print = tokenizer.decode(inp[0])
cur_text = tokenizer.decode(inp[0],skip_special_tokens=True)
gen_texts.append(cur_text)
if best_clip_score < current_clip_score.cpu().item():
best_clip_score = current_clip_score.cpu().item()
best_caption = cur_text
logger.info(f"iter {iter_num + 1}, clip score {current_clip_score:.3f}: "+for_print)
gen_texts.append(best_caption)
clip_score_sequence.append(best_clip_score)
return gen_texts, clip_score_sequence
def span_generation(model, clip, tokenizer,image_instance,token_mask, prompt, logger,
max_len=15, top_k=0,temperature=None, alpha=0.7,beta=1,
max_iters=20,batch_size=1,verbose=True):
""" Generate multiple words at a time (span generation), in L->R order """
seed_len = len(prompt.split())+1
span_len = 2
batch = get_init_text(tokenizer,prompt, max_len, batch_size)
image_embeds = clip.compute_image_representation_from_image_instance(image_instance)
clip_score_sequence = []
best_clip_score = 0
inp = torch.tensor(batch).to(image_embeds.device)
gen_texts = []
for iter_num in range(max_iters):
for span_start in range(0,max_len,span_len):
span_end = min(span_start+span_len,max_len)
for jj in range(batch_size):
inp[jj][seed_len + span_start: seed_len + span_end] = tokenizer.mask_token_id
out = model(inp).logits
for ii in range(span_start,span_end):
token_mask = update_token_mask(tokenizer, token_mask, max_len, ii)
inp_ = inp.clone().detach()
probs, idxs = generate_caption_step(out, gen_idx=seed_len + ii, mask=token_mask, top_k=top_k,
temperature=temperature)
for jj in range(batch_size):
topk_inp = inp_.repeat(top_k, 1)
idxs_ = (idxs[jj] * token_mask[0][idxs[jj]]).long()
topk_inp[:, ii + seed_len] = idxs_
batch_text_list = tokenizer.batch_decode(topk_inp, skip_special_tokens=True)
clip_score, clip_ref = clip.compute_image_text_similarity_via_raw_text(image_embeds, batch_text_list)
final_score = alpha * probs + beta * clip_score
best_clip_id = final_score.argmax()
inp[jj][seed_len + ii] = idxs_[best_clip_id]
current_clip_score = clip_ref[jj][best_clip_id]
clip_score_sequence.append(current_clip_score.cpu().item())
if verbose and np.mod(iter_num + 1, 1) == 0:
for_print = tokenizer.decode(inp[0])
cur_text = tokenizer.decode(inp[0],skip_special_tokens=True)
if best_clip_score < current_clip_score.cpu().item():
best_clip_score = current_clip_score.cpu().item()
best_caption = cur_text
gen_texts.append(cur_text)
logger.info(f"iter {iter_num + 1}, clip score {current_clip_score:.3f}: "+ for_print)
gen_texts.append(best_caption)
clip_score_sequence.append(best_clip_score)
return gen_texts, clip_score_sequence
def random_generation(model, clip, tokenizer,image_instance,token_mask, prompt, logger,
max_len=15, top_k=0, temperature=None,alpha=0.7,beta=2,
max_iters=300,print_every=10,batch_size=1,
verbose=True):
""" Generate for one random position at a timestep"""
seed_len = len(prompt.split())+1
batch = get_init_text(tokenizer, prompt, max_len, batch_size)
image_embeds = clip.compute_image_representation_from_image_instance(image_instance)
clip_score_sequence = []
best_clip_score = 0
inp = torch.tensor(batch).to(image_embeds.device)
gen_texts = []
for ii in range(max_iters):
kk = np.random.randint(0, max_len)
token_mask = update_token_mask(tokenizer, token_mask, max_len, kk)
for jj in range(batch_size):
inp[jj][seed_len + kk] = tokenizer.mask_token_id
inp_ = inp.clone().detach()
out = model(inp).logits
probs, idxs = generate_caption_step(out,gen_idx=seed_len + kk,mask=token_mask, top_k=top_k, temperature=temperature)
for jj in range(batch_size):
topk_inp = inp_.repeat(top_k, 1)
topk_inp[:, kk + seed_len] = (idxs[jj] * token_mask[0][idxs[jj]]).long()
batch_text_list = tokenizer.batch_decode(topk_inp, skip_special_tokens=True)
clip_score, clip_ref = clip.compute_image_text_similarity_via_raw_text(image_embeds, batch_text_list)
final_score = alpha * probs + beta * clip_score
best_clip_id = final_score.argmax()
inp[jj][seed_len + kk] = idxs[jj][best_clip_id]
current_clip_score = clip_ref[jj][best_clip_id]
clip_score_sequence.append(current_clip_score.cpu().item())
if best_clip_score < current_clip_score.cpu().item():
best_clip_score = current_clip_score.cpu().item()
best_caption = tokenizer.decode(inp[0], skip_special_tokens=True)
if verbose and np.mod(ii + 1, print_every) == 0:
for_print = tokenizer.decode(inp[0])
logger.info(f"iter {ii + 1}, clip score {current_clip_score:.3f}: "+for_print)
cur_text = tokenizer.decode(inp[0], skip_special_tokens=True)
gen_texts.append(cur_text)
gen_texts.append(best_caption)
clip_score_sequence.append(best_clip_score)
return gen_texts, clip_score_sequence
def parallel_generation(model, clip, tokenizer,image_instance,token_mask, prompt, logger,
max_len=15, top_k=0, temperature=None, alpha=0.1, beta=1,
max_iters=300,batch_size=1,print_every=1, verbose=True):
""" Generate for all positions at a time step """
seed_len = len(prompt.split())+1
batch = get_init_text(tokenizer,prompt, max_len, batch_size)
image_embeds = clip.compute_image_representation_from_image_instance(image_instance)
clip_score_sequence = []
inp = torch.tensor(batch).to(image_embeds.device)
gen_texts = []
best_clip_score = 0
for ii in range(max_iters):
inp_ = inp.clone().detach()
out = model(inp).logits
for kk in range(max_len):
probs, idxs = generate_caption_step(out, gen_idx=seed_len + kk,mask=token_mask, top_k=top_k, temperature=temperature)
for jj in range(batch_size):
topk_inp = inp_.repeat(top_k, 1)
topk_inp[:, ii + seed_len] = (idxs[jj] * token_mask[0][idxs[jj]]).long()
batch_text_list = tokenizer.batch_decode(topk_inp, skip_special_tokens=True)
clip_score,clip_ref = clip.compute_image_text_similarity_via_raw_text(image_embeds, batch_text_list)
final_score = alpha * probs + beta * clip_score
best_clip_id = final_score.argmax()
inp[jj][seed_len + kk] = idxs[jj][best_clip_id]
current_clip_score = clip_ref[jj][best_clip_id]
clip_score_sequence.append(current_clip_score.cpu().item())
if verbose and np.mod(ii, 1) == 0:
logger.info(f"iter{ii + 1}, clip score {current_clip_score:.3f}: " + tokenizer.decode(inp[0]))
cur_text = tokenizer.decode(inp[0], skip_special_tokens=True)
if best_clip_score < current_clip_score.cpu().item():
best_clip_score = current_clip_score.cpu().item()
best_caption = cur_text
gen_texts.append(cur_text)
gen_texts.append(best_caption)
clip_score_sequence.append(best_clip_score)
return gen_texts, clip_score_sequence
def generate_caption(model, clip, tokenizer,image_instance,token_mask,logger,
prompt="", batch_size=1, max_len=15,
top_k=100, temperature=1.0, max_iter=500,alpha=0.7,beta=1,
generate_order="sequential"):
# main generation functions to call
start_time = time.time()
if generate_order=="sequential":
generate_texts, clip_scores = sequential_generation(model, clip, tokenizer, image_instance, token_mask, prompt, logger,
batch_size=batch_size, max_len=max_len, top_k=top_k,
alpha=alpha,beta=beta,temperature=temperature,
max_iters=max_iter)
elif generate_order=="shuffle":
# max_iter = 15
generate_texts, clip_scores = shuffle_generation(model, clip, tokenizer,image_instance,token_mask,prompt, logger,
batch_size=batch_size, max_len=max_len, top_k=top_k,
alpha=alpha,beta=beta,temperature=temperature,max_iters=max_iter)
elif generate_order=="random":
max_iter *= max_len
print_every = max_len
generate_texts, clip_scores = random_generation(model, clip, tokenizer,image_instance,token_mask,prompt,logger,
max_len=max_len, top_k=top_k,alpha=alpha,beta=beta,print_every=print_every,
temperature=temperature, max_iters=max_iter,verbose=True)
elif generate_order=="span":
max_iter = max_iter
generate_texts, clip_scores = span_generation(model, clip, tokenizer, image_instance, token_mask, prompt, logger,
batch_size=batch_size, max_len=max_len, top_k=top_k,
alpha=alpha,beta=beta,temperature=temperature, max_iters=max_iter)
elif generate_order=="parallel":
generate_texts, clip_scores = parallel_generation(model, clip, tokenizer,image_instance,token_mask,prompt, logger,
max_len=max_len, temperature=temperature,top_k=top_k,alpha=alpha,beta=beta,
max_iters=max_iter,verbose=True)
logger.info("Finished in %.3fs" % (time.time() - start_time))
logger.info(f"final caption: {generate_texts[-2]}")
logger.info(f"best caption: {generate_texts[-1]}")
return generate_texts, clip_scores |