File size: 6,211 Bytes
ce922b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# 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
from typing import Optional
from argparse import Namespace

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

from tasks.ofa_task import OFATask, OFAConfig
from data.mm_data.refcoco_dataset import RefcocoDataset
from data.file_dataset import FileDataset

logger = logging.getLogger(__name__)


@dataclass
class RefcocoConfig(OFAConfig):
    # options for reporting BLEU during validation
    eval_acc: bool = field(
        default=False, metadata={"help": "evaluation with BLEU scores"}
    )
    eval_args: Optional[str] = field(
        default='{}',
        metadata={
            "help": 'generation args for BLUE or CIDEr scoring, e.g., \'{"beam": 4, "lenpen": 0.6}\', as JSON string'
        },
    )
    eval_print_samples: bool = field(
        default=False, metadata={"help": "print sample generations during validation"}
    )

    max_image_size: int = field(
        default=512, metadata={"help": "max image size for normalization"}
    )
    scst: bool = field(
        default=False, metadata={"help": "Self-critical sequence training"}
    )
    scst_args: str = field(
        default='{}',
        metadata={
            "help": 'generation args for Self-critical sequence training, as JSON string'
        },
    )


@register_task("refcoco", dataclass=RefcocoConfig)
class RefcocoTask(OFATask):
    def __init__(self, cfg: RefcocoConfig, src_dict, tgt_dict):
        super().__init__(cfg, src_dict, tgt_dict)

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

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

        self.datasets[split] = RefcocoDataset(
            split,
            dataset,
            self.bpe,
            self.src_dict,
            self.tgt_dict,
            max_src_length=self.cfg.max_src_length,
            max_tgt_length=self.cfg.max_tgt_length,
            patch_image_size=self.cfg.patch_image_size,
            imagenet_default_mean_and_std=self.cfg.imagenet_default_mean_and_std,
            num_bins=self.cfg.num_bins,
            max_image_size=self.cfg.max_image_size
        )

    def build_model(self, cfg):
        model = super().build_model(cfg)
        if self.cfg.eval_acc:
            gen_args = json.loads(self.cfg.eval_args)
            self.sequence_generator = self.build_generator(
                [model], Namespace(**gen_args)
            )
        if self.cfg.scst:
            scst_args = json.loads(self.cfg.scst_args)
            self.scst_generator = self.build_generator(
                [model], Namespace(**scst_args)
            )

        return model

    def _calculate_ap_score(self, 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()

    def valid_step(self, sample, model, criterion):
        loss, sample_size, logging_output = criterion(model, sample)

        model.eval()
        if self.cfg.eval_acc:
            hyps, refs = self._inference(self.sequence_generator, sample, model)
            hyps = hyps / (self.cfg.num_bins - 1) * self.cfg.max_image_size
            refs = refs / (self.cfg.num_bins - 1) * self.cfg.max_image_size
            hyps[:, ::2] /= sample['w_resize_ratios'].unsqueeze(1)
            hyps[:, 1::2] /= sample['h_resize_ratios'].unsqueeze(1)
            refs[:, ::2] /= sample['w_resize_ratios'].unsqueeze(1)
            refs[:, 1::2] /= sample['h_resize_ratios'].unsqueeze(1)

            # scores = self._calculate_ap_score(hyps, refs)
            scores = self._calculate_ap_score(hyps, sample['region_coords'].float())
            logging_output["_score_sum"] = scores.sum().item()
            logging_output["_score_cnt"] = scores.size(0)

        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["_score_sum"].sum / meters["_score_cnt"].sum
            score = score if isinstance(score, float) else score.item()
            return round(score, 4)

        if sum_logs("_score_cnt") > 0:
            metrics.log_scalar("_score_sum", sum_logs("_score_sum"))
            metrics.log_scalar("_score_cnt", sum_logs("_score_cnt"))
            metrics.log_derived("score", compute_score)

    def _inference(self, generator, sample, model):
        gen_out = self.inference_step(generator, [model], sample)
        hyps, refs = [], []
        for i in range(len(gen_out)):
            hyps.append(gen_out[i][0]["tokens"][:-1] - len(self.src_dict) + self.cfg.num_bins)
            refs.append(sample["target"][i][:-1] - len(self.src_dict) + self.cfg.num_bins)
        if self.cfg.eval_print_samples:
            logger.info("example hypothesis: ", hyps[0])
            logger.info("example reference: ", refs[0])

        return torch.stack(hyps, dim=0), torch.stack(refs, dim=0)