Metis/val.py at main · YouAreSpecialToMe/Metis · 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
import torch
import seaborn as sns
import matplotlib.pyplot as plt
from tqdm import tqdm
from sklearn.metrics import confusion_matrix, f1_score, precision_score, recall_score


def val(model, intent_dataloader, epoch, mode='DEV', logger=None, config=None, i2in=None,
        criterion=torch.nn.CrossEntropyLoss()):
    assert mode in ['TRAIN', 'DEV', 'TEST', 'INIT']

    avg_loss = 0
    acc = 0

    if intent_dataloader is None:
        return None, None, None, None, None

    model.eval()
    pbar_dev = tqdm(intent_dataloader)
    pbar_dev.set_description("VAL {}".format(mode))
    all_pred = []
    all_label = []
    with torch.no_grad():
        for batch in pbar_dev:

            if config.model_type == 'Onehot':
                x = batch['x']
            else:
                x_forward = batch['x_forward']
                x_backward = batch['x_backward']

            label = batch['i'].view(-1)
            lengths = batch['l']

            if torch.cuda.is_available():
                if config.model_type == 'Onehot':
                    x = x.cuda()
                else:
                    x_forward = x_forward.cuda()
                    x_backward = x_backward.cuda()

                lengths = lengths.cuda()
                label = label.cuda()

            if config.bidirection:
                out = model(x_forward, x_backward, lengths)
            else:
                if config.model_type == 'Onehot':
                    out = model(x, lengths)
                else:
                    out = model(x_forward, lengths)

            # print("out: ", out)
            # print("x: ",x)
            # print("label: ", label)
            # return 0

            loss = criterion(out, label)
            avg_loss += loss.item()
            acc += (out.argmax(1) == label).sum().item()
            all_pred += list(out.argmax(1).cpu().numpy())
            all_label += list(label.cpu().numpy())

            pbar_dev.set_postfix_str("{} - total right: {}, total entropy loss: {}".format(mode, acc, loss))

    acc = acc / len(intent_dataloader.dataset)
    avg_loss = avg_loss / len(intent_dataloader.dataset)
    p_micro, r_micro = acc, acc
    if 'SMS' in config.dataset:
        p_micro = precision_score(all_label, all_pred, average='binary', pos_label=1)
        r_micro = recall_score(all_label, all_pred, average='binary', pos_label=1)

    print('Dataset Len: {}'.format(len(intent_dataloader.dataset)))
    print("{} Epoch: {} | ACC: {}, LOSS: {}, P: {}, R: {}".format(mode, epoch, acc, avg_loss, p_micro, r_micro))
    if logger:
        logger.add(
            "{} Epoch: {} | ACC: {}, LOSS: {}, P: {}, R: {}".format(mode, epoch, acc, avg_loss, p_micro, r_micro))

    if config.only_probe:
        confusion_mat = confusion_matrix(all_label, all_pred, labels=[i for i in range(len(i2in))])
        labels = [i2in[i] for i in range(len(i2in))]
        fig = plt.figure()
        fig.set_size_inches(18, 18)
        cmap = sns.cubehelix_palette(8, start=2, rot=0, dark=0, light=.95, reverse=False)
        g = sns.heatmap(confusion_mat, annot=True, cmap=cmap, linewidths=1,
                        linecolor='gray', xticklabels=labels, yticklabels=labels, )
        plt.show()

    return acc, avg_loss, p_micro, r_micro


def val_marry(model, intent_dataloader, epoch, mode='DEV', config=None, logger=None,
              criterion=torch.nn.CrossEntropyLoss()):
    # assert mode in ['DEV', 'TEST']

    avg_loss = 0
    acc = 0

    if not intent_dataloader:
        return None, None, None, None

    model.eval()
    pbar_dev = tqdm(intent_dataloader)
    pbar_dev.set_description("VAL {}".format(mode))
    all_pred = []
    all_label = []
    with torch.no_grad():
        for batch in pbar_dev:

            x = batch['x']
            label = batch['i'].view(-1)
            lengths = batch['l']
            re_tag = batch['re']

            if torch.cuda.is_available():
                x = x.cuda()
                lengths = lengths.cuda()
                label = label.cuda()
                re_tag = re_tag.cuda()

            out = model(x, lengths, re_tag)
            loss = criterion(out, label)
            avg_loss += loss.item()
            acc += (out.argmax(1) == label).sum().item()

            all_pred += list(out.argmax(1).cpu().numpy())
            all_label += list(label.cpu().numpy())

            pbar_dev.set_postfix_str("{} - total right: {}, total entropy loss: {}".format(mode, acc, loss))

    acc = acc / len(intent_dataloader.dataset)
    avg_loss = avg_loss / len(intent_dataloader.dataset)
    p_micro, r_micro = acc, acc
    if config.dataset == 'SMS':
        p_micro = precision_score(all_label, all_pred, average='binary', pos_label=1)
        r_micro = recall_score(all_label, all_pred, average='binary', pos_label=1)

    print("{} Epoch: {} | ACC: {}, LOSS: {}, P: {}, R: {}".format(mode, epoch, acc, avg_loss, p_micro, r_micro))
    if logger:
        logger.add(
            "{} Epoch: {} | ACC: {}, LOSS: {}, P: {}, R: {}".format(mode, epoch, acc, avg_loss, p_micro, r_micro))

    return acc, avg_loss, p_micro, r_micro