2024-iACP-GCR/train_func.py at main · mldlproject/2024-iACP-GCR · 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
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
import numpy as np
import time
import torch
from sklearn.metrics import precision_recall_curve
from sklearn.metrics import confusion_matrix
from sklearn.metrics import roc_auc_score
from sklearn import metrics
from sklearn import metrics
from sklearn.metrics import matthews_corrcoef, accuracy_score, f1_score, precision_score, recall_score, confusion_matrix
from sklearn.metrics import roc_auc_score, average_precision_score, cohen_kappa_score, balanced_accuracy_score


#
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')


def get_loss_task(criterion, num_classes_tasks, probs, labels, device):
    total_loss = 0
    for t_id in range(num_classes_tasks):
        y_pred = probs[t_id] # output of each task
        y_label = labels[:, t_id:t_id+1].squeeze() # label of task
        validId = np.where((y_label.cpu().numpy() == 0) | (y_label.cpu().numpy() == 1))[0] # index tại những dữ liệu tồn tại label
        if len(validId) == 0:
            continue
        if y_label.dim() == 0:
            y_label = y_label.unsqueeze(0)

        y_pred = y_pred[torch.tensor(validId).to(device)]
        y_label = y_label[torch.tensor(validId).to(device)]

        loss = criterion(y_pred.view(-1).squeeze(), y_label.squeeze())
        total_loss += loss
    total_loss = total_loss/num_classes_tasks
    return total_loss

def get_prob_task(num_classes_tasks, probs, labels, device):
    prob_list = []
    label_list = []
    for t_id in range(num_classes_tasks):
        y_pred = probs[t_id] # output of each task
        y_label = labels[:, t_id:t_id+1].squeeze() # label of task
        validId = np.where((y_label.cpu().numpy() == 0) | (y_label.cpu().numpy() == 1))[0] # index tại những dữ liệu tồn tại label
        if len(validId) == 0:
            continue
        if y_label.dim() == 0:
            y_label = y_label.unsqueeze(0)

        y_pred = y_pred[torch.tensor(validId).to(device)]
        y_label = y_label[torch.tensor(validId).to(device)]

        prob_list.append(y_pred.detach().cpu().view_as(y_label).numpy()) # lưu giá trị dự đoán
        label_list.append(y_label.detach().cpu().numpy()) # lưu label dữ liệu

    return prob_list, label_list

def get_performace(y_label_list, y_pred_list, tasks):
    trn_roc =  np.array([roc_auc_score(y_label_list[i], y_pred_list[i]) for i in range(len(tasks))])
    trn_prc =  np.array([metrics.auc(precision_recall_curve(y_label_list[i], y_pred_list[i])[1],
                        precision_recall_curve(y_label_list[i], y_pred_list[i])[0]) for i in range(len(tasks))])
    #---------------------------------------------
    #--------------------------------------------
    predicted_labels = [[] for i in range(len(tasks))]
    for i in range(len(tasks)):
        for prob in y_pred_list[i]:
            predicted_labels[i].append(np.round(prob))

    # acc = accuracy_score(labels, predicted_labels)
    trn_acc = np.array([accuracy_score(y_label_list[i], predicted_labels[i]) for i in range(len(tasks))])
    trn_ba  = np.array([balanced_accuracy_score(y_label_list[i], predicted_labels[i]) for i in range(len(tasks))])
    trn_mcc = np.array([matthews_corrcoef(y_label_list[i], predicted_labels[i]) for i in range(len(tasks))])
    trn_ck  =  np.array([cohen_kappa_score(y_label_list[i], predicted_labels[i]) for i in range(len(tasks))])
    trn_sensitivity, trn_specificity, trn_precision, trn_f1 = [], [], [], []
    for i in range(len(tasks)):
        tn, fp, fn, tp = confusion_matrix(y_label_list[i], predicted_labels[i]).ravel()
        sensitivity = tp / (tp + fn)
        specificity = tn / (tn + fp)
        precision   = tp / (tp + fp)
        f1 = 2*precision*sensitivity / (precision + sensitivity)
        trn_sensitivity.append(sensitivity)
        trn_specificity.append(specificity)
        trn_precision.append(precision)
        trn_f1.append(f1)

    trn_sensitivity, trn_specificity, trn_precision, trn_f1 = np.array(trn_sensitivity), np.array(trn_specificity), np.array(trn_precision), np.array(trn_f1)

    perform = [trn_roc, trn_prc, trn_acc, trn_ba, trn_mcc, trn_ck, trn_sensitivity, trn_specificity, trn_precision, trn_f1]
    return perform

##########################################################################################
###                  DEFINE TRAINING, VALIDATION, AND TEST FUNCTION                    ###
##########################################################################################
# Training Function
def train_funct(epoch, model, optimizer, criterion, tasks, train_loader):
    model.train()
    train_loss = 0
    start_time = time.time()
    for batch_idx, batch in enumerate(train_loader):
        data       = batch.to(device)
        labels     = batch.y.to(device).float() # Chứa label của 12 task

        outputs = model(data)
        #-------------------
        optimizer.zero_grad()
        #-------------------
        avg_loss = get_loss_task(criterion, len(tasks), outputs, labels, device)

        avg_loss.backward()
        train_loss += avg_loss.item()*len(data) #(loss.item is the average loss of training batch)
        optimizer.step()

    #-------------------
    print('====> Epoch: {}, training time {},  Average Train Loss: {:.4f}'.format(epoch, time.time() - start_time, train_loss / len(train_loader)))
    train_loss = (train_loss / len(train_loader.dataset) )

    return train_loss

##########################################################################################
# Validation Function
def validate(epoch, model, criterion, tasks, val_loader):
    model.eval()
    validation_loss = 0
    pred_prob = []
    y_pred_list = {}
    y_label_list = {}
    for i in range(len(tasks)):
        y_label_list[i] = []
        y_pred_list[i] = []
    with torch.no_grad():
        for batch_idx, batch in enumerate(val_loader):
            data       = batch.to(device)
            labels     = batch.y.to(device).float() # Chứa label của 12 task

            outputs = model(data)

            avg_loss = get_loss_task(criterion, len(tasks), outputs, labels, device)
            validation_loss += avg_loss.item()*len(data) #(loss.item is the average loss of training batch)

            pred_list, label_list = get_prob_task(len(tasks), outputs, labels, device) # Chưa predict của 5 task

            # print(len(label_list))
            for i in range(len(tasks)):
                if label_list[i] != 'None':
                    y_label_list[i].extend(label_list[i])
                    y_pred_list[i].extend(pred_list[i])

    print('====> Epoch: {} Average Validation Loss: {:.4f}'.format(epoch, validation_loss / len(val_loader)))
    validation_loss = (validation_loss / len(val_loader.dataset) )
    perform = get_performace(y_label_list, y_pred_list, tasks)

    return validation_loss, perform

# ##########################################################################################
# Test Function
def test(current_iter, model, tasks, test_loader):
    model.eval()
    y_pred_list = {}
    y_label_list = {}
    for i in range(len(tasks)):
        y_label_list[i] = []
        y_pred_list[i] = []
    with torch.no_grad():
        for batch_idx, batch in enumerate(test_loader):
            data       = batch.to(device)
            labels     = batch.y.to(device).float() # Chứa label của 12 task
            outputs = model(data)
            pred_list, label_list = get_prob_task(len(tasks), outputs, labels, device) # Chưa predict của 5 task

            # print(len(label_list))
            for i in range(len(tasks)):
                if label_list[i] != 'None':

                    y_label_list[i].extend(label_list[i])
                    y_pred_list[i].extend(pred_list[i])
    perform = get_performace(y_label_list, y_pred_list, tasks)
    print("Performance of model at epoch {} on test dataset".format(current_iter))
    print("AUC of {} task: {}".format(len(tasks),perform[0]))
    print("PR_AUC of {} task: {}".format(len(tasks),perform[1]))
    print("######################################################")
    return perform