-
Notifications
You must be signed in to change notification settings - Fork 0
/
classification.py
211 lines (157 loc) · 6.51 KB
/
classification.py
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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
#coding=utf-8
# @Time : 20-12-24下午3:11
# @Author : Honglian WANG
import os
import sys
from sklearn.metrics import classification_report
import numpy as np
from numpy import Inf
from torch import nn
import torch
import torch.optim as optim
from self_attention import MultiHeadedAttention
from config import config
from load_dataset import load_data
from tools import mini_batch, train_test_split
import torch.nn.functional as F
import sklearn
from matplotlib import pyplot as plt
# import warnings
# warnings.filterwarnings('ignore')
class GestureClassification(torch.nn.Module):
def __init__(self,fea_dim):
super().__init__()
self.fea_dim = fea_dim
self.lstm = torch.nn.LSTM(fea_dim, config.rnn_h_dim,
bidirectional=False, batch_first=True)
self.attention = MultiHeadedAttention(config.n_head,
config.rnn_h_dim, config.att_dim, config.att_drop)
self.mlp1 = torch.nn.Linear(config.att_dim, 16)
self.mlp2 = torch.nn.Linear(16, 8)
self.softmax = nn.Softmax(dim=1)
# batch normalization
self.dense_bn1 = nn.BatchNorm1d(config.att_dim)
self.dense_bn2 = nn.BatchNorm1d(16)
self.dense_bn3 = nn.BatchNorm1d(8)
def forward(self, x, real_len, mask):
''' feed data and give output'''
hidden = self.lstm(x)[0] # [batch_size, seq_len, embed_dim]
query = self.tensor_indexing(hidden, real_len) #[batch_size, 1, embed_dim]
att_out = self.attention(query,hidden,hidden,mask)
att_out = att_out + query # residuel connection [b,1,att_dim]
att_out = torch.squeeze(att_out,dim=1) # [b,1,att_dim]->[b,att_dim]
if x.shape[0] == 1: # batch_norm requires more than 1 data on each channel
# att_out.shape = [1, seq_len]
x = att_out / torch.sum(att_out)
p = self.mlp1(x)
p = F.relu(p/torch.sum(p))
p = self.mlp2(p)
p = F.relu(p/torch.sum(p))
prediction = self.softmax(p)
else:
x = self.dense_bn1(att_out)
p = self.mlp1(x)
p = F.relu(self.dense_bn2(p))
p = self.mlp2(p)
p = F.relu(self.dense_bn3(p))
prediction = self.softmax(p)
return prediction
def tensor_indexing(self, tensor, index):
d1, d2, d3 = tensor.shape
index = index - 1 # index should be len of sequence -1
index1 = index.reshape((d1, 1))
index2 = torch.from_numpy(index1).long()
index3 = index2.repeat(1, d3)
index4 = torch.unsqueeze(index3, 1)
return torch.gather(tensor, 1, index4.to(device))
def batch_training(model, train_X, train_S, train_Y, train_M, optimizer, lr_scheduler):
acc_arr = []
recall_arr = []
loss_arr = []
for epoch in range(config.EPOCH):
total_loss = 0
minibatch = mini_batch(train_X, train_S, train_Y, train_M)
cnt = 0
for (x, s, label, mask) in minibatch:
# perform prediction
prediction = model(x,s,mask) # [num_seq * num_label]
task_loss = F.cross_entropy(prediction, label)
optimizer.zero_grad()
task_loss.backward()
optimizer.step()
lr_scheduler.step()
total_loss += task_loss.item()
ranked_prediction = torch.argmax(prediction, dim=1)
prediction_c = ranked_prediction.cpu().data.numpy()
cnt += 1
if cnt == 1:
PRE = prediction_c
LAB = label.cpu().data.numpy()
if cnt > 1:
PRE = np.concatenate((PRE,prediction_c), axis=0)
LAB = np.concatenate((LAB,label.cpu().data.numpy()), axis=0)
print ('total loss at epoch %d is : %f' %(epoch, total_loss))
if epoch % 10 == 0:
torch.save(model.state_dict(), 'model/'+
str(epoch) + '.pkl')
loss_arr.append(total_loss)
acc, recall = metric(LAB, PRE)
acc_arr.append(acc)
recall_arr.append(recall)
plot_acc(acc_arr, recall_arr)
def plot_acc(acc, recall):
x = list(np.arange(len(acc)))
l1 = plt.plot(x, acc, 'r--', label='accuracy')
l2 = plt.plot(x, recall, 'g--', label='recall')
plt.plot(x, acc, 'r-', x, recall, 'g-')
plt.title('The accuracy and recall curve')
plt.xlabel('epoch')
plt.ylabel('value')
plt.legend()
plt.show()
def metric(y_true, y_pred):
acc = sklearn.metrics.accuracy_score(y_true, y_pred)
recall = sklearn.metrics.recall_score(y_true, y_pred, average='macro')
return acc, recall
def Prediction(model, test_X, test_L, test_Y,test_M):
with torch.no_grad():
minibatch = mini_batch(test_X, test_L, test_Y,test_M)
cnt = 0
PRE = []
LAB = []
for (x, s, label, mask) in minibatch:
prediction = model(x,s, mask) # [num_seq * num_label]
ranked_prediction = torch.argmax(prediction, dim=1)
prediction = ranked_prediction.cpu().data.numpy()
cnt += 1
if cnt == 1:
PRE = prediction
LAB = label
if cnt > 1:
PRE = np.concatenate((PRE,prediction), axis=0)
LAB = np.concatenate((LAB,label), axis=0)
print(classification_report(LAB, PRE))
if __name__ == '__main__':
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# device = 'cpu'
Data, Mask, Seq_len, Label = load_data(config.file_path)
print('Data.shape', Data.shape)
fea_dim = Data.shape[-1]
[train_X,train_L,train_Y,train_M], [test_X, test_L, test_Y,test_M] = train_test_split(Data, Seq_len, Label, Mask)
print ('train_X.shape', train_X.shape)
print ('testX.shape', test_X.shape)
model = GestureClassification(fea_dim)
model.to(device)
optimizer = optim.Adam(model.parameters(), lr=config.lr, weight_decay=config.lambda_l2_reg)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=4)
train_X = train_X.to(device).float()
train_M = train_M.to(device).long()
train_Y = torch.tensor(train_Y, device=device)
batch_training(model, train_X, train_L, train_Y, train_M, optimizer, scheduler)
test_X = test_X.to(device).float()
test_M = test_M.to(device).long()
# model = GestureClassification(fea_dim)
# model.load_state_dict(torch.load('model/90.pkl'))
# model.to(device)
Prediction(model, test_X, test_L, test_Y, test_M)