DeepKWS/train.py at master · WarmCongee/DeepKWS · 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

from random import shuffle
from network import *
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision.models import resnet18
from data_set import KWSDataSet

Labels = {"null": 0, "Hello": 1, "xiaogua": 2, "nihao": 3, "xiaoyi": 4, "jixu": 5, "tingzhi": 6, "bofang":7}
key_class = ("Hello", "xiaogua", "nihao", "xiaoyi", "jixu", "tingzhi", "bofang", "null")
train_dataset_path = "/home/disk1/user5/wyz/DataSet/TrainSet"
valid_dataset_path = "/home/disk1/user5/wyz/DataSet/ValidSet"
test_dataset_path = "/home/disk1/user5/wyz/DataSet/TestSet"


def generate_batch_data(batch):
    batch_X = []
    batch_Y = []
    for index, data in enumerate(batch):
        batch_datas = []
        batch_labels = []
        frame_list_2d = torch.split(data, 1, dim=0)
        for frame_item in frame_list_2d:
            batch_item_list = torch.split(frame_item, 40, dim=1)
            batch_datas.append(batch_item_list[0])
            batch_labels.append(batch_item_list[1])
        temp_x, temp_y = combine_frames_as_frame(batch_datas, batch_labels)
        if index==0:
            batch_X = temp_x
            batch_Y = temp_y
        else:
            batch_X = torch.cat((batch_X,temp_x),0)
            batch_Y = torch.cat((batch_Y,temp_y),0)
    return batch_X, batch_Y


def combine_frames_as_frame(datas, labels):
    re_datas = torch.zeros(2, 41, 40)
    re_labels = torch.zeros(2)
    for index,data in enumerate(datas):
        if(index >= 30 and index < len(datas)-10):
            sum_tensor = datas[index-30]
            for j in range(index-29,index+11):
                sum_tensor = torch.cat((sum_tensor,datas[j]), 0)
            re_datas = torch.cat((re_datas, torch.unsqueeze(sum_tensor, 0)), 0)
            re_labels = torch.cat((re_labels, torch.squeeze(labels[index], 0)), 0)
    return re_datas, re_labels

fbanks_batch_size = 4
device = torch.device('cuda:2' if torch.cuda.is_available() else 'cpu')
print('using device:', device)


train_set = KWSDataSet(train_dataset_path)
train_loader = torch.utils.data.DataLoader(train_set, batch_size=fbanks_batch_size, collate_fn=generate_batch_data,
                                            shuffle=True, num_workers=16, pin_memory=True)
valid_set = KWSDataSet(valid_dataset_path)
valid_loader = torch.utils.data.DataLoader(valid_set, batch_size=fbanks_batch_size, collate_fn=generate_batch_data,
                                            shuffle=True, num_workers=16, pin_memory=True)


net = resnet18()
num_ftrs = net.fc.in_features
net.fc = nn.Linear(num_ftrs, 8)
net.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
net.to(device)

criterion = nn.CrossEntropyLoss().to(device)
optimizer = optim.SGD(net.parameters(), lr=0.003)
StepLR = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.80)
loss_list = []
log_steps = 3000

def valid():

    count_valid = 0
    loss_sum_valid = 0.0
    correct = 0
    net.eval()
    with torch.no_grad():
        for index, data in enumerate(valid_loader):

            inputs = data[0].to(device)
            labels =data[1].to(device, dtype=torch.int64)
            inputs = torch.unsqueeze(inputs, 1)
            outputs = net(inputs)
            loss = criterion(outputs, labels)
            loss_sum_valid = loss_sum_valid + loss
            correct += (outputs.argmax(1)==labels).type(torch.float).sum().item()
            count_valid += labels.shape[0]

            del inputs, labels
            if index % 3000 == 2999:
                print(f'average valid loss: {loss_sum_valid / (index+1):.7f}')
                print(f'acc: {correct / (count_valid):.7f}')

def train():
    for epoch in range(20):
        net.train()
        print("learning rate of epoch-%d: %f" % (epoch, optimizer.param_groups[0]['lr']))
        loss_epoch_sum = 0.0
        loss_temp_sum = 0.0
        for index, data in enumerate(train_loader):
            inputs = data[0].to(device)
            labels = data[1].to(device, dtype=torch.int64)
            inputs = torch.unsqueeze(inputs, 1)
            optimizer.zero_grad()
            outputs = net(inputs)

            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()
            del inputs, labels
            loss_temp_sum += loss

            if index % 50 == 49:
                loss_list.append(loss)
            if index % log_steps == (log_steps-1):
                loss_epoch_sum += loss_temp_sum
                print(f'{epoch}-{index+1} average loss: {loss_temp_sum / log_steps:.7f}')
                loss_temp_sum = 0.0


        print(f'[{epoch + 1}] average train loss: {loss_epoch_sum / (index+1):.7f}')
        StepLR.step()


        if epoch % 4 == 3:
            torch.save(net.state_dict(), 'deep_kws-data2-net3-'+str(epoch)+'.pth')
            torch.save(loss_list, 'loss_list-data2-net3-'+str(epoch)+'.pth')
        print("-------valid--------")
        valid()
        print("--------------------")


if __name__ == '__main__':
    train()

    torch.save(net.state_dict(), 'deep_kws-net3.pth')
    print('Finished Training')