ms-code-100/supervisor_train.py at main · 2023-MindSpore-1/ms-code-100 · 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
import time
import numpy as np
import mindspore
from mindspore import nn, optim
import copy

from model import *
from dataloader_d import *

import warnings
warnings.filterwarnings("ignore")

class ModelSupervisor_train(nn.Module):
    def __init__(self, channel, channel_n, kernel_size, timeslot, scaler_n, resnet_n, step, feature, poi, device=0):
        super(ModelSupervisor_train, self).__init__()
        self.device= device

        if feature is not None:
            self.feature = feature
            if self.feature.max() > 1:
                self.feature = self.feature / np.max(self.feature)
        self.poi = poi
        if self.poi.max() > 1:
            self.poi = self.poi / np.max(self.poi)

        self.cmpnet = CMPNet(channel, channel, kernel_size, timeslot, self.feature, self.poi, self.device)
        self.sr = SRNet(channel, channel_n, scaler_n, resnet_n, step, poi.shape[-1])

    def train(self, iteration, cmp_lr, sr_lr, best_rmse, pretrain_d, train_d, pretrain_model, pretrain_model_save, train_model_save):
        # restore cmpnet
        net1 = self.cmpnet.to(device=self.device)
        net1.load_state_dict(mindspore.load_checkpoint(pretrain_model))
        net2 = self.sr.to(device=self.device)

        optimizer = optim.Adam([
            {'params': net2.parameters(), 'lr': sr_lr, 'betas': (0.9, 0.999)},
            {'params': net1.parameters(), 'lr': cmp_lr, 'betas': (0.9, 0.999)},
        ])

        for epoch in range(iteration):
            start = time.time()
            loss_RMSE = []
            loss_MAE = []
            # training
            for i, (pretrain, train) in enumerate(zip(pretrain_d[0], train_d[0])):
                # pretrain[3]: X, Y, Feature; train[3]: X, Y, Feature
                mask_train = copy.deepcopy(pretrain[0].to(self.device))
                mask_train[mask_train != 0] = 1
                label_pre = pretrain[1].to(self.device)

                data_tra = train[0].to(self.device)
                label_tra = train[1].to(self.device)
                # 判断是否存在 external features
                if len(train) == 3:
                    ext_tra = train[2].to(self.device)
                else:
                    ext_tra = None

                output1 = net1(data_tra, mask_train)
                output2 = net2(output1, ext_tra)

                loss_mae_pre = self.mae_loss(output1, label_pre)
                loss_mae = self.mae_loss(output2, label_tra)
                loss_rmse = self.rmse_loss(output2, label_tra)
                loss_MAE.append(loss_mae.item())
                loss_RMSE.append(loss_rmse.item())

                # multi
                total_mae = 0.0001 * loss_mae_pre + 0.1 * loss_mae
                optimizer.zero_grad()
                total_mae.backward(retain_graph=True)
                optimizer.step()

            t = time.time() - start
            result = 'epoch: {}, time: {:.4f}, train_mae: {:.4f}, train_rmse: {:.4f} '.format(epoch, t, np.mean(loss_MAE),
                                                                                            np.sqrt(np.mean(loss_RMSE)))
            print(result)

            # evaluating
            if (epoch+1) % 10 == 0:
                jdg, best_rmse = self.evaluate(net1, net2, best_rmse, train_d[1])
                if jdg == True:
                    print("best model!")
                    best_model_2 = net2
                    best_model_1 = net1
                # testing
                self.test(best_model_1, best_model_2, train_d[2])

            if epoch % 20 == 0 and epoch != 0:
               sr_lr /= 2
               optimizer = optim.Adam(net2.parameters(), lr=sr_lr)

        net1.load_state_dict(mindspore.load_checkpoint(pretrain_model_save))
        net2.load_state_dict(mindspore.load_checkpoint(train_model_save))

    def evaluate(self, net1, net2, b_r, train_val):
        jdg = False
        start = time.time()
        loss_RMSE = []
        loss_MAE = []
        for i, sample in enumerate(train_val):
            mask_train = copy.deepcopy(sample[0].to(self.device))
            mask_train[mask_train != 0] = 1

            data = sample[0].to(self.device)
            label = sample[1].to(self.device)
            if len(sample) == 3:
                ext = ext.to(self.device)
            else:
                ext = None

            output1 = net1(data, mask_train)
            output2 = net2(output1, ext)

            loss_mae = self.mae_loss(output2, label)
            loss_rmse = self.rmse_loss(output2, label)
            loss_MAE.append(loss_mae.item())
            loss_RMSE.append(loss_rmse.item())

        t = time.time() - start
        result = 'evaluating -- time: {:.4f}, val_mae: {:.4f}, val_rmse: {:.4f} '.format(t, np.mean(loss_MAE),
                                                                                           np.sqrt(np.mean(loss_RMSE)))
        print(result)
        if np.sqrt(np.mean(loss_RMSE)) < b_r:
            jdg = True
            b_r = np.sqrt(np.mean(loss_RMSE))
        return jdg, b_r

    def test(self, net1, best_model, train_test):
        start = time.time()
        loss_RMSE = []
        loss_MAE = []
        for i, sample in enumerate(train_test):
            mask_train = copy.deepcopy(sample[0].to(self.device))
            mask_train[mask_train != 0] = 1

            data = sample[1].to(self.device)
            label = sample[1].to(self.device)
            if len(sample) == 3:
                ext = ext.to(self.device)
            else:
                ext = None

            output1 = net1(data, mask_train)
            output2 = best_model(output1, ext)

            loss_mae = self.mae_loss(output2, label)
            loss_rmse = self.rmse_loss(output2, label)
            loss_MAE.append(loss_mae.item())
            loss_RMSE.append(loss_rmse.item())

        t = time.time() - start
        result = 'testing -- time: {:.4f}, test_mae: {:.4f}, test_rmse: {:.4f} '.format(t, np.mean(loss_MAE),
                                                                                        np.sqrt(np.mean(loss_RMSE)))
        print(result)

    def mae_loss(self, y_pred, y_true):
        mask = (y_true != 0).to(self.device).float()
        mask /= mask.mean()
        loss = mindspore.ops.abs(y_pred - y_true)
        loss = loss * mask
        loss[loss != loss] = 0
        return loss.mean()

    def rmse_loss(self, y_pred, y_true):
        mask = (y_true != 0).to(self.device).float()
        mask /= mask.mean()
        loss = mindspore.ops.Pow((y_pred - y_true), 2)
        loss = loss * mask
        loss[loss != loss] = 0
        return loss.mean()