File size: 9,625 Bytes
0d735a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

from dataclasses import dataclass, field
import json
import logging
import os
import math
import pickle
from typing import Optional
from argparse import Namespace
from data.file_dataset import FileDataset

import torch
from fairseq import metrics
from fairseq.tasks import register_task

from models import search
from data.mm_data.vqa_gen_dataset import VqaGenDataset
from data import data_utils
from tasks.ofa_task import OFAConfig, OFATask
from utils.trie import Trie

logger = logging.getLogger(__name__)


@dataclass
class VqaGenConfig(OFAConfig):
    max_object_length: int = field(
        default=30, metadata={"help": "the maximum object sequence length"}
    )    
    ans2label_dict: Optional[str] = field(
        default='{"no": 0, "yes":1}',
        metadata={"help": 'answer to label dict'},
    )
    ans2label_file: Optional[str] = field(
        default=None,
        metadata={"help": "path to load ans2label file"},
    )

    add_object: bool = field(
        default=False,
        metadata={"help": "add object to encoder"},
    )
    valid_batch_size: int = field(
        default=20,
        metadata={"help": "valid batch size per step"},
    )
    prompt_type: Optional[str] = field(
        default=None,
        metadata={"help": "prompt_type"},
    )
    uses_ema: Optional[bool] = field(
        default=False,
        metadata={"help": "whether to use ema"},
    )


@register_task("vqa_gen", dataclass=VqaGenConfig)
class VqaGenTask(OFATask):
    def __init__(self, cfg: VqaGenConfig, src_dict, tgt_dict):
        super().__init__(cfg, src_dict, tgt_dict)

        self.ans2label_dict = None
        if self.cfg.ans2label_file is not None:
            self.ans2label_dict = pickle.load(open(self.cfg.ans2label_file, "rb"))
        else:
            self.ans2label_dict = json.loads(self.cfg.ans2label_dict)

        self.uses_ema = self.cfg.uses_ema

    def load_dataset(self, split, epoch=1, combine=False, **kwargs):
        paths = self.cfg.data.split(',')
        assert len(paths) > 0

        if split == 'train':
            table_path = paths[(epoch - 1) % (len(paths) - 1)]
        else:
            table_path = paths[-1]
        dataset = FileDataset(table_path, self.cfg.selected_cols)

        self.datasets[split] = VqaGenDataset(
            split,
            dataset,
            self.bpe,
            self.src_dict,
            self.tgt_dict,
            max_src_length=self.cfg.max_src_length,
            max_object_length=self.cfg.max_object_length,
            max_tgt_length=self.cfg.max_tgt_length,
            patch_image_size=self.cfg.patch_image_size,
            add_object=self.cfg.add_object,
            constraint_trie=self.constraint_trie,
            imagenet_default_mean_and_std=self.cfg.imagenet_default_mean_and_std,
            prompt_type=self.cfg.prompt_type
        )

    def build_model(self, cfg):
        model = super().build_model(cfg)
        answer_item_list = []
        self.index2ans = {}
        self.constraint_trie = Trie(self.tgt_dict.eos())
        for i, answer in enumerate(self.ans2label_dict.keys()):
            answer_item = self.tgt_dict.encode_line(
                line=self.bpe.encode(' ' + answer),
                add_if_not_exist=False,
                append_eos=False
            ).long()
            answer_item_list.append(answer_item)
            self.index2ans[i] = answer
            self.constraint_trie.insert([self.tgt_dict.bos()] + answer_item.tolist() + [self.tgt_dict.eos()])

        constraint_mask_list = []
        for answer_item in answer_item_list:
            constraint_mask = torch.zeros((len(answer_item)+1, len(self.tgt_dict))).bool()
            for i in range(len(answer_item)+1):
                constraint_prefix_token = [self.src_dict.bos()] + answer_item[:i].tolist()
                constraint_nodes = self.constraint_trie.get_next_layer(constraint_prefix_token)
                constraint_mask[i][constraint_nodes] = True
            constraint_mask_list.append(constraint_mask)

        self.valid_answers_list = []
        self.valid_constraint_masks_list = []
        for i in range(0, len(answer_item_list), self.cfg.valid_batch_size):
            self.valid_answers_list += [answer_item_list[i:i+self.cfg.valid_batch_size]]
            self.valid_constraint_masks_list += [constraint_mask_list[i:i+self.cfg.valid_batch_size]]

        return model

    def build_generator(
        self, models, args, seq_gen_cls=None, extra_gen_cls_kwargs=None, prefix_allowed_tokens_fn=None,
    ):
        seq_generator = super().build_generator(models, args, seq_gen_cls, extra_gen_cls_kwargs, prefix_allowed_tokens_fn)
        seq_generator.constraint_trie = self.constraint_trie

        return seq_generator

    def valid_step(self, sample, model, criterion, **extra_kwargs):
        loss, sample_size, logging_output = super().valid_step(sample, model, criterion)

        if self.uses_ema:
            assert 'ema_model' in extra_kwargs and extra_kwargs['ema_model'] is not None
        if self.uses_ema:
            eval_model = extra_kwargs['ema_model']
        else:
            eval_model = model

        eval_model.eval()
        with torch.no_grad():
            encoder_out = eval_model.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([self.src_dict.eos()])
            pad = self.src_dict.pad()
            valid_result = []
            for valid_answers, valid_constraint_masks in zip(self.valid_answers_list, self.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, left_pad=False).to(device)
                valid_prev_output = data_utils.collate_tokens(valid_prev_items, pad_idx=pad, left_pad=False).to(device)
                valid_constraint_masks = data_utils.collate_tokens(valid_constraint_mask_items, pad_idx=pad, left_pad=False).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 = eval_model.decoder(valid_prev_output, encoder_out=new_encoder_out)
                decoder_out[0].masked_fill_(~valid_constraint_masks, -math.inf)
                lprobs = eval_model.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(self.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 = [self.index2ans[predict_index] for predict_index in predicts]
        scores = [ref_dict.get(hyp, 0) for ref_dict, hyp in zip(sample['ref_dict'], hyps)]
        logging_output["_vqa_score_sum"] = sum(scores)
        logging_output["_vqa_cnt"] = len(scores)

        return loss, sample_size, logging_output

    def reduce_metrics(self, logging_outputs, criterion):
        super().reduce_metrics(logging_outputs, criterion)

        def sum_logs(key):
            import torch
            result = sum(log.get(key, 0) for log in logging_outputs)
            if torch.is_tensor(result):
                result = result.cpu()
            return result

        def compute_score(meters):
            score = meters["_vqa_score_sum"].sum / meters["_vqa_cnt"].sum
            score = score if isinstance(score, float) else score.item()
            return round(score, 4)

        if sum_logs("_vqa_cnt") > 0:
            metrics.log_scalar("_vqa_score_sum", sum_logs("_vqa_score_sum"))
            metrics.log_scalar("_vqa_cnt", sum_logs("_vqa_cnt"))
            metrics.log_derived("vqa_score", compute_score)