File size: 7,476 Bytes
7734d5b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
# Copyright (c) Megvii, Inc. and its affiliates.

from loguru import logger
from tqdm import tqdm

import torch

from yolox.utils import (
    gather,
    is_main_process,
    postprocess,
    synchronize,
    time_synchronized,
    xyxy2xywh
)

import contextlib
import io
import itertools
import json
import tempfile
import time


class COCOEvaluator:
    """
    COCO AP Evaluation class.  All the data in the val2017 dataset are processed
    and evaluated by COCO API.
    """

    def __init__(
        self, dataloader, img_size, confthre, nmsthre, num_classes, testdev=False
    ):
        """
        Args:
            dataloader (Dataloader): evaluate dataloader.
            img_size (int): image size after preprocess. images are resized
                to squares whose shape is (img_size, img_size).
            confthre (float): confidence threshold ranging from 0 to 1, which
                is defined in the config file.
            nmsthre (float): IoU threshold of non-max supression ranging from 0 to 1.
        """
        self.dataloader = dataloader
        self.img_size = img_size
        self.confthre = confthre
        self.nmsthre = nmsthre
        self.num_classes = num_classes
        self.testdev = testdev

    def evaluate(
        self,
        model,
        distributed=False,
        half=False,
        trt_file=None,
        decoder=None,
        test_size=None,
    ):
        """
        COCO average precision (AP) Evaluation. Iterate inference on the test dataset
        and the results are evaluated by COCO API.

        NOTE: This function will change training mode to False, please save states if needed.

        Args:
            model : model to evaluate.

        Returns:
            ap50_95 (float) : COCO AP of IoU=50:95
            ap50 (float) : COCO AP of IoU=50
            summary (sr): summary info of evaluation.
        """
        # TODO half to amp_test
        tensor_type = torch.cuda.HalfTensor if half else torch.cuda.FloatTensor
        model = model.eval()
        if half:
            model = model.half()
        ids = []
        data_list = []
        progress_bar = tqdm if is_main_process() else iter

        inference_time = 0
        nms_time = 0
        n_samples = len(self.dataloader) - 1

        if trt_file is not None:
            from torch2trt import TRTModule

            model_trt = TRTModule()
            model_trt.load_state_dict(torch.load(trt_file))

            x = torch.ones(1, 3, test_size[0], test_size[1]).cuda()
            model(x)
            model = model_trt

        for cur_iter, (imgs, _, info_imgs, ids) in enumerate(
            progress_bar(self.dataloader)
        ):
            with torch.no_grad():
                imgs = imgs.type(tensor_type)

                # skip the the last iters since batchsize might be not enough for batch inference
                is_time_record = cur_iter < len(self.dataloader) - 1
                if is_time_record:
                    start = time.time()

                outputs = model(imgs)
                if decoder is not None:
                    outputs = decoder(outputs, dtype=outputs.type())

                if is_time_record:
                    infer_end = time_synchronized()
                    inference_time += infer_end - start

                outputs = postprocess(
                    outputs, self.num_classes, self.confthre, self.nmsthre
                )
                if is_time_record:
                    nms_end = time_synchronized()
                    nms_time += nms_end - infer_end

            data_list.extend(self.convert_to_coco_format(outputs, info_imgs, ids))

        statistics = torch.cuda.FloatTensor([inference_time, nms_time, n_samples])
        if distributed:
            data_list = gather(data_list, dst=0)
            data_list = list(itertools.chain(*data_list))
            torch.distributed.reduce(statistics, dst=0)

        eval_results = self.evaluate_prediction(data_list, statistics)
        synchronize()
        return eval_results

    def convert_to_coco_format(self, outputs, info_imgs, ids):
        data_list = []
        for (output, img_h, img_w, img_id) in zip(
            outputs, info_imgs[0], info_imgs[1], ids
        ):
            if output is None:
                continue
            output = output.cpu()

            bboxes = output[:, 0:4]

            # preprocessing: resize
            scale = min(
                self.img_size[0] / float(img_h), self.img_size[1] / float(img_w)
            )
            bboxes /= scale
            bboxes = xyxy2xywh(bboxes)

            cls = output[:, 6]
            scores = output[:, 4] * output[:, 5]
            for ind in range(bboxes.shape[0]):
                label = self.dataloader.dataset.class_ids[int(cls[ind])]
                pred_data = {
                    "image_id": int(img_id),
                    "category_id": label,
                    "bbox": bboxes[ind].numpy().tolist(),
                    "score": scores[ind].numpy().item(),
                    "segmentation": [],
                }  # COCO json format
                data_list.append(pred_data)
        return data_list

    def evaluate_prediction(self, data_dict, statistics):
        if not is_main_process():
            return 0, 0, None

        logger.info("Evaluate in main process...")

        annType = ["segm", "bbox", "keypoints"]

        inference_time = statistics[0].item()
        nms_time = statistics[1].item()
        n_samples = statistics[2].item()

        a_infer_time = 1000 * inference_time / (n_samples * self.dataloader.batch_size)
        a_nms_time = 1000 * nms_time / (n_samples * self.dataloader.batch_size)

        time_info = ", ".join(
            [
                "Average {} time: {:.2f} ms".format(k, v)
                for k, v in zip(
                    ["forward", "NMS", "inference"],
                    [a_infer_time, a_nms_time, (a_infer_time + a_nms_time)],
                )
            ]
        )

        info = time_info + "\n"

        # Evaluate the Dt (detection) json comparing with the ground truth
        if len(data_dict) > 0:
            cocoGt = self.dataloader.dataset.coco
            # TODO: since pycocotools can't process dict in py36, write data to json file.
            if self.testdev:
                json.dump(data_dict, open("./yolox_testdev_2017.json", "w"))
                cocoDt = cocoGt.loadRes("./yolox_testdev_2017.json")
            else:
                _, tmp = tempfile.mkstemp()
                json.dump(data_dict, open(tmp, "w"))
                cocoDt = cocoGt.loadRes(tmp)
            '''
            try:
                from yolox.layers import COCOeval_opt as COCOeval
            except ImportError:
                from pycocotools import cocoeval as COCOeval
                logger.warning("Use standard COCOeval.")
            '''
            #from pycocotools.cocoeval import COCOeval
            from yolox.layers import COCOeval_opt as COCOeval
            cocoEval = COCOeval(cocoGt, cocoDt, annType[1])
            cocoEval.evaluate()
            cocoEval.accumulate()
            redirect_string = io.StringIO()
            with contextlib.redirect_stdout(redirect_string):
                cocoEval.summarize()
            info += redirect_string.getvalue()
            return cocoEval.stats[0], cocoEval.stats[1], info
        else:
            return 0, 0, info