Spaces:
Runtime error
Runtime error
File size: 13,295 Bytes
c9bb3f2 |
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 |
import string
import math
import json
from itertools import chain
import os
import torch
import torch.distributed as dist
from data import data_utils
def get_symbols_to_strip_from_output(generator):
if hasattr(generator, "symbols_to_strip_from_output"):
return generator.symbols_to_strip_from_output
else:
return {generator.bos, generator.eos}
def decode_fn(x, tgt_dict, bpe, generator, tokenizer=None):
x = tgt_dict.string(x.int().cpu(), extra_symbols_to_ignore=get_symbols_to_strip_from_output(generator))
if bpe is not None:
x = bpe.decode(x)
if tokenizer is not None:
x = tokenizer.decode(x)
return x
def eval_caption(task, generator, models, sample, **kwargs):
transtab = str.maketrans({key: None for key in string.punctuation})
hypos = task.inference_step(generator, models, sample)
results = []
for i, sample_id in enumerate(sample["id"].tolist()):
detok_hypo_str = decode_fn(hypos[i][0]["tokens"], task.tgt_dict, task.bpe, generator)
results.append({"image_id": str(sample_id), "caption": detok_hypo_str.translate(transtab).strip()})
return results, None
def eval_vqa_gen(task, generator, models, sample, **kwargs):
if kwargs['beam_search_vqa_eval']:
hypos = task.inference_step(generator, models, sample, prefix_tokens=sample['prefix_tokens'])
results = []
for i, sample_id in enumerate(sample["id"].tolist()):
prefix_len = sample['prefix_tokens'][i].ne(1).sum().item()
detok_hypo_str = decode_fn(hypos[i][0]["tokens"][prefix_len:], task.tgt_dict, task.bpe, generator)
results.append({"question_id": int(sample_id), "answer": detok_hypo_str.strip()})
scores = [ref_dict.get(result['answer'], 0) for ref_dict, result in zip(sample['ref_dict'], results)]
return results, scores
encoder_out = models[0].encoder(
sample["net_input"]["src_tokens"],
src_lengths=sample["net_input"]["src_lengths"],
patch_images=sample["net_input"]["patch_images"],
patch_masks=sample["net_input"]["patch_masks"]
)
device = sample["net_input"]["src_tokens"].device
eos_item = torch.tensor([task.src_dict.eos()])
pad = task.src_dict.pad()
valid_result = []
for valid_answers, valid_constraint_masks in zip(task.valid_answers_list, task.valid_constraint_masks_list):
valid_size = len(valid_answers)
valid_tgt_items = [
torch.cat([torch.tensor(decoder_prompt[1:]), valid_answer, eos_item])
for decoder_prompt in sample["decoder_prompts"] for valid_answer in valid_answers
]
valid_prev_items = [
torch.cat([torch.tensor(decoder_prompt), valid_answer])
for decoder_prompt in sample["decoder_prompts"] for valid_answer in valid_answers
]
valid_constraint_mask_items = [
torch.cat(
[torch.zeros(len(decoder_prompt) - 1, valid_constraint_mask.size(1)).bool(), valid_constraint_mask],
dim=0
)
for decoder_prompt in sample["decoder_prompts"] for valid_constraint_mask in valid_constraint_masks
]
valid_tgt = data_utils.collate_tokens(valid_tgt_items, pad_idx=pad).to(device)
valid_prev_output = data_utils.collate_tokens(valid_prev_items, pad_idx=pad).to(device)
valid_constraint_masks = data_utils.collate_tokens(valid_constraint_mask_items, pad_idx=pad).to(device)
new_encoder_out = {}
new_encoder_out["encoder_out"] = [
encoder_out["encoder_out"][0].repeat_interleave(valid_size, dim=1)
]
new_encoder_out["encoder_padding_mask"] = [
encoder_out["encoder_padding_mask"][0].repeat_interleave(valid_size, dim=0)
]
new_encoder_out["position_embeddings"] = [
encoder_out["position_embeddings"][0].repeat_interleave(valid_size, dim=0)
]
decoder_out = models[0].decoder(valid_prev_output, encoder_out=new_encoder_out)
decoder_out[0].masked_fill_(~valid_constraint_masks, -math.inf)
lprobs = models[0].get_normalized_probs(decoder_out, log_probs=True)
scores = lprobs.gather(dim=-1, index=valid_tgt.unsqueeze(-1)).squeeze(-1)
scores = scores.masked_fill(valid_tgt.eq(task.tgt_dict.pad()), 0)
scores = scores.masked_fill((~valid_constraint_masks).all(2), 0)
scores = scores.sum(1)
scores = scores.view(-1, valid_size)
valid_result.append(scores)
valid_result = torch.cat(valid_result, dim=-1)
predicts = valid_result.argmax(1).tolist()
hyps = [task.index2ans[predict_index] for predict_index in predicts]
results = [{"question_id": int(id), "answer": hyp} for id, hyp in zip(sample["id"].tolist(), hyps)]
scores = [ref_dict.get(hyp, 0) for ref_dict, hyp in zip(sample['ref_dict'], hyps)]
return results, scores
def eval_refcoco(task, generator, models, sample, **kwargs):
def _calculate_ap_score(hyps, refs, thresh=0.5):
interacts = torch.cat(
[torch.where(hyps[:, :2] < refs[:, :2], refs[:, :2], hyps[:, :2]),
torch.where(hyps[:, 2:] < refs[:, 2:], hyps[:, 2:], refs[:, 2:])],
dim=1
)
area_predictions = (hyps[:, 2] - hyps[:, 0]) * (hyps[:, 3] - hyps[:, 1])
area_targets = (refs[:, 2] - refs[:, 0]) * (refs[:, 3] - refs[:, 1])
interacts_w = interacts[:, 2] - interacts[:, 0]
interacts_h = interacts[:, 3] - interacts[:, 1]
area_interacts = interacts_w * interacts_h
ious = area_interacts / (area_predictions + area_targets - area_interacts + 1e-6)
return ((ious >= thresh) & (interacts_w > 0) & (interacts_h > 0)).float()
gen_out = task.inference_step(generator, models, sample)
hyps = []
for i in range(len(gen_out)):
hyps.append(gen_out[i][0]["tokens"][:-1] - len(task.src_dict) + task.cfg.num_bins)
hyps = torch.stack(hyps, dim=0)
hyps = hyps / (task.cfg.num_bins - 1) * task.cfg.max_image_size
hyps[:, ::2] /= sample['w_resize_ratios'].unsqueeze(1)
hyps[:, 1::2] /= sample['h_resize_ratios'].unsqueeze(1)
results = [
{"uniq_id": sample_id,
"box": [hyps[i][0].item(), hyps[i][1].item(), hyps[i][2].item(), hyps[i][3].item()]}
for i, sample_id in enumerate(sample["id"].tolist())
]
scores = _calculate_ap_score(hyps, sample['region_coords'].float())
return results, scores
def eval_snli_ve(task, generator, models, sample, **kwargs):
encoder_out = models[0].encoder(
sample["net_input"]["src_tokens"],
src_lengths=sample["net_input"]["src_lengths"],
patch_images=sample["net_input"]["patch_images"],
patch_masks=sample["net_input"]["patch_masks"]
)
device = sample["net_input"]["src_tokens"].device
eos_item = torch.tensor([task.src_dict.eos()])
pad = task.src_dict.pad()
valid_result = []
for valid_answers, valid_constraint_masks in zip(task.valid_answers_list, task.valid_constraint_masks_list):
valid_size = len(valid_answers)
valid_tgt_items = [
torch.cat([torch.tensor(decoder_prompt[1:]), valid_answer, eos_item])
for decoder_prompt in sample["decoder_prompts"] for valid_answer in valid_answers
]
valid_prev_items = [
torch.cat([torch.tensor(decoder_prompt), valid_answer])
for decoder_prompt in sample["decoder_prompts"] for valid_answer in valid_answers
]
valid_constraint_mask_items = [
torch.cat(
[torch.zeros(len(decoder_prompt) - 1, valid_constraint_mask.size(1)).bool(), valid_constraint_mask],
dim=0
)
for decoder_prompt in sample["decoder_prompts"] for valid_constraint_mask in valid_constraint_masks
]
valid_tgt = data_utils.collate_tokens(valid_tgt_items, pad_idx=pad).to(device)
valid_prev_output = data_utils.collate_tokens(valid_prev_items, pad_idx=pad).to(device)
valid_constraint_masks = data_utils.collate_tokens(valid_constraint_mask_items, pad_idx=pad).to(device)
new_encoder_out = {}
new_encoder_out["encoder_out"] = [
encoder_out["encoder_out"][0].repeat_interleave(valid_size, dim=1)
]
new_encoder_out["encoder_padding_mask"] = [
encoder_out["encoder_padding_mask"][0].repeat_interleave(valid_size, dim=0)
]
new_encoder_out["position_embeddings"] = [
encoder_out["position_embeddings"][0].repeat_interleave(valid_size, dim=0)
]
decoder_out = models[0].decoder(valid_prev_output, encoder_out=new_encoder_out)
decoder_out[0].masked_fill_(~valid_constraint_masks, -math.inf)
lprobs = models[0].get_normalized_probs(decoder_out, log_probs=True)
scores = lprobs.gather(dim=-1, index=valid_tgt.unsqueeze(-1)).squeeze(-1)
scores = scores.masked_fill(valid_tgt.eq(task.tgt_dict.pad()), 0)
scores = scores.masked_fill((~valid_constraint_masks).all(2), 0)
scores = scores.sum(1)
scores = scores.view(-1, valid_size)
valid_result.append(scores)
valid_result = torch.cat(valid_result, dim=-1)
predicts = valid_result.argmax(1).tolist()
hyps = [task.index2ans[predict_index] for predict_index in predicts]
results = [{"uniq_id": id, "answer": hyp} for id, hyp in zip(sample["id"].tolist(), hyps)]
scores = [ref_dict.get(hyp, 0) for ref_dict, hyp in zip(sample['ref_dict'], hyps)]
return results, scores
def eval_image_gen(task, generator, models, sample, **kwargs):
hypos, _ = task.inference_image(generator, sample, models)
tokens = sample['net_input']['src_tokens'][0].view(-1).tolist()
caption = task.bpe.decode(task.tgt_dict.string([token for token in tokens if token >= 4]))[
38:].replace('/', '')
text_similarity_score, indices = task.compute_text_similarity(hypos, caption,
sample['net_input']['src_tokens'].device)
results = []
for i, indice in enumerate(indices):
results.append({"sample_id": str(sample["id"][0]), "score": text_similarity_score[i], "image": hypos[indice]})
scores = [max(text_similarity_score).item()]
return results, scores
def eval_glue(task, generator, models, sample, **kwargs):
net_output = models[0](**sample["net_input"])
net_output[0].masked_fill_(~sample["constraint_masks"], -math.inf)
last_token_ids = sample["net_input"]["prev_output_tokens"].ne(task.src_dict.pad()).sum(1, keepdim=True) - 1
logits = net_output[0].gather(1, last_token_ids.unsqueeze(2).expand(-1, -1, net_output[0].size(2)))
logits = logits.squeeze(1)
predicts = logits.argmax(1).tolist()
hyps = [task.bpe.decode(task.src_dict[predict]).strip() for predict in predicts]
results = [{"hyp": hyp, "ref": ref_dict.keys()[0]} for hyp, ref_dict in zip(hyps, sample['ref_dict'])]
return results, None
def eval_step(task, generator, models, sample, **kwargs):
if task.cfg._name == 'caption':
return eval_caption(task, generator, models, sample, **kwargs)
elif task.cfg._name == 'vqa_gen':
return eval_vqa_gen(task, generator, models, sample, **kwargs)
elif task.cfg._name == 'refcoco':
return eval_refcoco(task, generator, models, sample, **kwargs)
elif task.cfg._name == 'snli_ve':
return eval_snli_ve(task, generator, models, sample, **kwargs)
elif task.cfg._name == 'image_gen':
return eval_image_gen(task, generator, models, sample, **kwargs)
elif task.cfg._name in {'cola', 'mnli', 'mrpc', 'qnli', 'qqp', 'rte', 'sst2'}:
return eval_glue(task, generator, models, sample, **kwargs)
else:
raise NotImplementedError
def merge_results(task, cfg, logger, score_cnt, score_sum, results):
if task.cfg._name == 'image_gen':
if cfg.distributed_training.distributed_world_size > 1:
dist.all_reduce(score_sum.data)
dist.all_reduce(score_cnt.data)
if score_cnt.item() > 0:
logger.info("score_sum: {}, score_cnt: {}, score: {}".format(
score_sum, score_cnt, round(score_sum.item() / score_cnt.item(), 4)
))
else:
gather_results = None
if cfg.distributed_training.distributed_world_size > 1:
gather_results = [None for _ in range(dist.get_world_size())]
dist.all_gather_object(gather_results, results)
dist.all_reduce(score_sum.data)
dist.all_reduce(score_cnt.data)
if score_cnt.item() > 0:
logger.info("score_sum: {}, score_cnt: {}, score: {}".format(
score_sum, score_cnt, round(score_sum.item() / score_cnt.item(), 4)
))
if cfg.distributed_training.distributed_world_size == 1 or dist.get_rank() == 0:
os.makedirs(cfg.common_eval.results_path, exist_ok=True)
output_path = os.path.join(cfg.common_eval.results_path, "{}_predict.json".format(cfg.dataset.gen_subset))
gather_results = list(chain(*gather_results)) if gather_results is not None else results
with open(output_path, 'w') as fw:
json.dump(gather_results, fw)
|