AST/train_teacher.py at main · apple024/AST · 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
import time
import numpy as np
import torch
from os.path import join
from sklearn.metrics import roc_auc_score
from tqdm import tqdm

import config as c
from model import *
from utils import *


def train(train_loader, test_loader):
    model = Model()
    model.to(c.device)
    optimizer = torch.optim.Adam(model.net.parameters(), lr=c.lr, eps=1e-08, weight_decay=1e-5)

    mean_nll_obs = Score_Observer('AUROC mean over maps')
    max_nll_obs = Score_Observer('AUROC  max over maps')
    batch_limit_warning_printed = False

    start = time.time()
    train_log_step = 0
    test_log_step = 0

    for epoch in range(c.meta_epochs):
        # train some epochs
        model.train()
        if c.verbose:
            print(F'\nTrain epoch {epoch}')
        for sub_epoch in range(c.sub_epochs):
            train_loss = list()
            for i, data in enumerate(tqdm(train_loader, disable=c.hide_tqdm_bar)):
                optimizer.zero_grad()

                depth, fg, labels, image, features = data
                depth, fg, labels, image, features = to_device([depth, fg, labels, image, features])
                fg = dilation(fg, c.dilate_size) if c.dilate_mask else fg

                img_in = features if c.pre_extracted else image
                fg_down = downsampling(fg, (c.map_len, c.map_len), bin=False)
                z, jac = model(img_in, depth)

                loss = get_nf_loss(z, jac, fg_down)
                train_loss.append(t2np(loss))

                loss.backward()
                optimizer.step()

                if c.sub_epoch_batch_limit > 0 and i > c.sub_epoch_batch_limit:
                    if not batch_limit_warning_printed:
                        batch_limit_warning_printed = True
                        print(f"Batch limit of {c.sub_epoch_batch_limit} used to limit time per sub epoch.")
                    break

            mean_train_loss = np.mean(train_loss)
            if "mlflow_tracking_uri" in globals():
                mlflow.log_metric(f"{c.class_name}-teacher-train-loss", mean_train_loss, step=train_log_step)
                train_log_step += 1
            if c.verbose and sub_epoch % c.sub_epoch_log_interval == 0:  # and epoch == 0:
                print('Epoch: {:d}.{:d} \t train loss: {:.4f} \t ({:d} s runtime)'.format(epoch, sub_epoch, mean_train_loss, int(time.time() - start)))

        # evaluate
        model.eval()
        if c.verbose:
            print('\nCompute loss and scores on test set:')
        test_loss = list()
        test_labels = list()
        img_nll = list()
        max_nlls = list()
        with torch.no_grad():
            for i, data in enumerate(tqdm(test_loader, disable=c.hide_tqdm_bar)):
                depth, fg, labels, image, features = data
                depth, fg, image, features = to_device([depth, fg, image, features])

                fg = dilation(fg, c.dilate_size) if c.dilate_mask else fg

                img_in = features if c.pre_extracted else image
                fg_down = downsampling(fg, (c.map_len, c.map_len), bin=False)
                z, jac = model(img_in, depth)
                loss = get_nf_loss(z, jac, fg_down, per_sample=True)
                nll = get_nf_loss(z, jac, fg_down, per_pixel=True)

                img_nll.append(t2np(loss))
                max_nlls.append(np.max(t2np(nll), axis=(-1, -2)))
                test_loss.append(loss.mean().item())
                test_labels.append(labels)

        img_nll = np.concatenate(img_nll)
        max_nlls = np.concatenate(max_nlls)
        test_loss = np.mean(np.array(test_loss))

        if "mlflow_tracking_uri" in globals():
            mlflow.log_metric(f"{c.class_name}-teacher-test-loss", test_loss, step=test_log_step)
            test_log_step += 1
        if c.verbose:
            print('Epoch: {:d} \t test_loss: {:.4f}'.format(epoch, test_loss))

        test_labels = np.concatenate(test_labels)
        is_anomaly = np.array([0 if l == 0 else 1 for l in test_labels])

        mean_nll_obs.update(roc_auc_score(is_anomaly, img_nll), epoch,
                            print_score=c.verbose or epoch == c.meta_epochs - 1)
        max_nll_obs.update(roc_auc_score(is_anomaly, max_nlls), epoch,
                           print_score=c.verbose or epoch == c.meta_epochs - 1)

    if c.save_model:
        save_weights(model, 'teacher')

    return mean_nll_obs, max_nll_obs


if __name__ == "__main__":
    train_dataset(train, img_aug=True)