File size: 12,512 Bytes
650c5f6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# ------------------------------------------------------------------------
# Modified from OFA (https://github.com/OFA-Sys/OFA)
# Copyright 2022 The OFA-Sys Team. 
# All rights reserved.
# This source code is licensed under the Apache 2.0 license 
# found in the LICENSE file in the root directory.
# ------------------------------------------------------------------------
# Modifications Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
# SPDX-License-Identifier: Apache-2.0

from dataclasses import dataclass, field
import logging
from typing import Optional
import os
import math
import numpy as np
import torch
from fairseq import metrics
from fairseq.tasks import register_task

from tasks.base_task import BaseTask, BaseConfig, load_bert_pretrained_weights
from data.refcoco_dataset import RefcocoDataset
from data.file_dataset import FileDataset

logger = logging.getLogger(__name__)


COO = 0  # <COO> class
SEP = 1  # <SEP> class
EOS = 2  # <EOS> class
bos_index = 0   # index for bos token
sep_index = 3  # index for separator token




@dataclass
class RefcocoConfig(BaseConfig):
    eval_acc: bool = field(
        default=False, metadata={"help": "evaluation with accuracy"}
    )
    eval_args: Optional[str] = field(
        default='{}',
        metadata={
            "help": 'generation args, e.g., \'{"beam": 4, "lenpen": 0.6}\', as JSON string'
        },
    )
    uses_ema: Optional[bool] = field(
        default=False,
        metadata={"help": "whether to use ema"},
    )
    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(BaseTask):
    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)
        bert_path = "../../pretrained_weights/bert-base-uncased-pytorch_model.bin"
        if os.path.exists(bert_path):
            load_bert_pretrained_weights(model.encoder.bert, bert_path)
        if cfg._name == 'polyformer_b':
            swin_path = "../../pretrained_weights/swin_base_patch4_window12_384_22k.pth"
        else:
            swin_path = "../../pretrained_weights/swin_large_patch4_window12_384_22k.pth"
        if os.path.exists(swin_path):
            model.encoder.embed_images.init_weights(pretrained=swin_path)
        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(sample, model)
            scores = self._calculate_ap_score(hyps.float(), refs.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, sample, model):
        hyps = self.inference_step(model, sample)
        refs = sample['region_coords'].float()
        hyps = hyps * self.cfg.max_image_size
        hyps[:, ::2] /= sample['w_resize_ratios'].unsqueeze(1)
        hyps[:, 1::2] /= sample['h_resize_ratios'].unsqueeze(1)
        return hyps, refs

    def inference_step(self, model, sample):
        with torch.no_grad():
            if isinstance(model, list):
                model = model[0]
            min_len = 6
            max_len = 210
            model.eval()
            img = sample["net_input"]["patch_images"]
            b = img.shape[0]
            prev_output_token_11 = [[bos_index] for _ in range(b)]
            prev_output_token_12 = [[bos_index] for _ in range(b)]
            prev_output_token_21 = [[bos_index] for _ in range(b)]
            prev_output_token_22 = [[bos_index] for _ in range(b)]
            delta_x1 = [[0] for _ in range(b)]
            delta_y1 = [[0] for _ in range(b)]
            delta_x2 = [[1] for _ in range(b)]
            delta_y2 = [[1] for _ in range(b)]

            gen_out = [[] for _ in range(b)]

            n_bins = self.cfg.num_bins

            unfinish_flag = np.ones(b)
            i = 0

            encoder_out = model.encoder(
                sample['net_input']['src_tokens'],
                src_lengths=sample['net_input']['src_lengths'],
                att_masks=sample['net_input']['att_masks'],
                patch_images=sample['net_input']['patch_images'],
                patch_masks=sample['net_input']['patch_masks'],
                token_embeddings=None,
                return_all_hiddens=False,
                sample_patch_num=None
            )

            while i < max_len and unfinish_flag.any():
                prev_output_tokens_11_tensor = torch.tensor(np.array(prev_output_token_11)).to(img.device).long()
                prev_output_tokens_12_tensor = torch.tensor(np.array(prev_output_token_12)).to(img.device).long()
                prev_output_tokens_21_tensor = torch.tensor(np.array(prev_output_token_21)).to(img.device).long()
                prev_output_tokens_22_tensor = torch.tensor(np.array(prev_output_token_22)).to(img.device).long()
                delta_x1_tensor = torch.tensor(np.array(delta_x1)).to(img.device)
                delta_x2_tensor = torch.tensor(np.array(delta_x2)).to(img.device)
                delta_y1_tensor = torch.tensor(np.array(delta_y1)).to(img.device)
                delta_y2_tensor = torch.tensor(np.array(delta_y2)).to(img.device)

                net_output = model.decoder(
                    prev_output_tokens_11_tensor,
                    prev_output_tokens_12_tensor,
                    prev_output_tokens_21_tensor,
                    prev_output_tokens_22_tensor,
                    delta_x1_tensor,
                    delta_y1_tensor,
                    delta_x2_tensor,
                    delta_y2_tensor,
                    code_masks=None,
                    encoder_out=encoder_out,
                    features_only=False,
                    alignment_layer=None,
                    alignment_heads=None,
                    src_lengths=sample['net_input']['src_lengths'],
                    return_all_hiddens=False
                )

                cls_output = net_output[0]
                cls_type = torch.argmax(cls_output, 2)
                reg_output = net_output[1]
                for j in range(b):
                    if unfinish_flag[j] == 1:  # prediction is not finished
                        cls_j = cls_type[j, i].item()
                        if cls_j == COO or (cls_j == EOS and i < min_len):
                            output_j_x, output_j_y = reg_output[j, i].cpu().numpy()
                            output_j_x = min(output_j_x, 1)
                            output_j_y = min(output_j_y, 1)

                            gen_out[j].extend([output_j_x, output_j_y])

                            output_j_x = output_j_x * (n_bins - 1)
                            output_j_y = output_j_y * (n_bins - 1)

                            output_j_x_floor = math.floor(output_j_x)
                            output_j_y_floor = math.floor(output_j_y)
                            output_j_x_ceil = math.ceil(output_j_x)
                            output_j_y_ceil = math.ceil(output_j_y)

                            # tokenization
                            prev_output_token_11[j].append(output_j_x_floor * n_bins + output_j_y_floor + 4)
                            prev_output_token_12[j].append(output_j_x_floor * n_bins + output_j_y_ceil + 4)
                            prev_output_token_21[j].append(output_j_x_ceil * n_bins + output_j_y_floor + 4)
                            prev_output_token_22[j].append(output_j_x_ceil * n_bins + output_j_y_ceil + 4)

                            delta_x = output_j_x - output_j_x_floor
                            delta_y = output_j_y - output_j_y_floor

                        elif cls_j == SEP:
                            gen_out[j].append(2)  # insert 2 indicating separator tokens
                            prev_output_token_11[j].append(sep_index)
                            prev_output_token_12[j].append(sep_index)
                            prev_output_token_21[j].append(sep_index)
                            prev_output_token_22[j].append(sep_index)
                            delta_x = 0
                            delta_y = 0
                        else:  # eos is predicted and i >= min_len
                            unfinish_flag[j] = 0
                            gen_out[j].append(-1)
                            prev_output_token_11[j].append(2)  # 2 is eos token
                            prev_output_token_12[j].append(2)  # 2 is eos token
                            prev_output_token_21[j].append(2)  # 2 is eos token
                            prev_output_token_22[j].append(2)  # 2 is eos token
                            delta_x = 0
                            delta_y = 0
                    else:  # prediction is finished
                        gen_out[j].append(-1)
                        prev_output_token_11[j].append(1)  # 1 is padding token
                        prev_output_token_12[j].append(1)
                        prev_output_token_21[j].append(1)
                        prev_output_token_22[j].append(1)
                        delta_x = 0
                        delta_y = 0
                    delta_x1[j].append(delta_x)
                    delta_y1[j].append(delta_y)
                    delta_x2[j].append(1 - delta_x)
                    delta_y2[j].append(1 - delta_y)
                i += 1
        print("inference step: ", i)
        return gen_out