STA/engine.py at main · Mark12Ding/STA · GitHub

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
import os
import numpy as np
import math
import sys
from typing import Iterable, Optional
import torch
# from mixup import Mixup
from timm.utils import accuracy, ModelEma
import utils
from scipy.special import softmax


@torch.no_grad()
def validation_one_epoch(data_loader, model, device):
    criterion = torch.nn.CrossEntropyLoss()

    metric_logger = utils.MetricLogger(delimiter="  ")
    header = 'Val:'

    # switch to evaluation mode
    model.eval()

    for batch in metric_logger.log_every(data_loader, 10, header):
        videos = batch[0]
        target = batch[1]
        videos = videos.to(device, non_blocking=True)
        target = target.to(device, non_blocking=True)

        # compute output
        with torch.cuda.amp.autocast():
            output = model(videos)
            loss = criterion(output, target)

        acc1, acc5 = accuracy(output, target, topk=(1, 5))

        batch_size = videos.shape[0]
        metric_logger.update(loss=loss.item())
        metric_logger.meters['acc1'].update(acc1.item(), n=batch_size)
        metric_logger.meters['acc5'].update(acc5.item(), n=batch_size)
    # gather the stats from all processes
    metric_logger.synchronize_between_processes()
    print('* Acc@1 {top1.global_avg:.3f} Acc@5 {top5.global_avg:.3f} loss {losses.global_avg:.3f}'
          .format(top1=metric_logger.acc1, top5=metric_logger.acc5, losses=metric_logger.loss))

    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}


@torch.no_grad()
def final_test(data_loader, model, device, file):
    criterion = torch.nn.CrossEntropyLoss()

    metric_logger = utils.MetricLogger(delimiter="  ")
    header = 'Test:'

    # switch to evaluation mode
    model.eval()
    final_result = []

    for batch in metric_logger.log_every(data_loader, 10, header):
        videos = batch[0]
        target = batch[1]
        ids = batch[2]
        chunk_nb = batch[3]
        split_nb = batch[4]
        videos = videos.to(device, non_blocking=True)
        target = target.to(device, non_blocking=True)

        # compute output
        with torch.cuda.amp.autocast():
            output = model(videos)
            loss = criterion(output, target)

        for i in range(output.size(0)):
            string = "{} {} {} {} {}\n".format(ids[i], \
                                                str(output.data[i].cpu().numpy().tolist()), \
                                                str(int(target[i].cpu().numpy())), \
                                                str(int(chunk_nb[i].cpu().numpy())), \
                                                str(int(split_nb[i].cpu().numpy())))
            final_result.append(string)

        acc1, acc5 = accuracy(output, target, topk=(1, 5))

        batch_size = videos.shape[0]
        metric_logger.update(loss=loss.item())
        metric_logger.meters['acc1'].update(acc1.item(), n=batch_size)
        metric_logger.meters['acc5'].update(acc5.item(), n=batch_size)

    if not os.path.exists(file):
        os.mknod(file)
    with open(file, 'w') as f:
        f.write("{}, {}\n".format(acc1, acc5))
        for line in final_result:
            f.write(line)
    # gather the stats from all processes
    metric_logger.synchronize_between_processes()
    print('* Acc@1 {top1.global_avg:.3f} Acc@5 {top5.global_avg:.3f} loss {losses.global_avg:.3f}'
          .format(top1=metric_logger.acc1, top5=metric_logger.acc5, losses=metric_logger.loss))

    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}


def merge(eval_path, num_tasks):
    dict_feats = {}
    dict_label = {}
    dict_pos = {}
    print("Reading individual output files")

    for x in range(num_tasks):
        file = os.path.join(eval_path, str(x) + '.txt')
        lines = open(file, 'r').readlines()[1:]
        for line in lines:
            line = line.strip()
            name = line.split('[')[0]
            label = line.split(']')[1].split(' ')[1]
            chunk_nb = line.split(']')[1].split(' ')[2]
            split_nb = line.split(']')[1].split(' ')[3]
            data = np.fromstring(line.split('[')[1].split(']')[0], dtype=np.float, sep=',')
            data = softmax(data)
            if not name in dict_feats:
                dict_feats[name] = []
                dict_label[name] = 0
                dict_pos[name] = []
            if chunk_nb + split_nb in dict_pos[name]:
                continue
            dict_feats[name].append(data)
            dict_pos[name].append(chunk_nb + split_nb)
            dict_label[name] = label
    print("Computing final results")

    input_lst = []
    print(len(dict_feats))
    for i, item in enumerate(dict_feats):
        input_lst.append([i, item, dict_feats[item], dict_label[item]])
    from multiprocessing import Pool
    p = Pool(64)
    ans = p.map(compute_video, input_lst)
    top1 = [x[1] for x in ans]
    top5 = [x[2] for x in ans]
    pred = [x[0] for x in ans]
    label = [x[3] for x in ans]
    final_top1 ,final_top5 = np.mean(top1), np.mean(top5)
    return final_top1*100 ,final_top5*100

def compute_video(lst):
    i, video_id, data, label = lst
    feat = [x for x in data]
    feat = np.mean(feat, axis=0)
    pred = np.argmax(feat)
    top1 = (int(pred) == int(label)) * 1.0
    top5 = (int(label) in np.argsort(-feat)[:5]) * 1.0
    return [pred, top1, top5, int(label)]