-
Notifications
You must be signed in to change notification settings - Fork 4
/
def_model.py
123 lines (115 loc) · 5.73 KB
/
def_model.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
import numpy as np
import tensorflow as tf
from tensorflow import keras
########################################## Building blocks ###############################################
class my_cnn2d():
def __init__(self,k_size=3,in_chs=1,n_filters=32):
# Determine initialization parameters
fan_in=k_size*k_size*in_chs
fan_out=k_size*k_size*n_filters
#Xavior initialization
self.w=tf.Variable(tf.random.normal([k_size, k_size, in_chs,n_filters],stddev=np.sqrt(2/(fan_in+fan_out)) ,dtype=tf.float32))
# self.w=tf.Variable(tf.random.uniform([k_size, k_size, in_chs,n_filters],minval=-np.sqrt(6/(fan_in+fan_out)),maxval=np.sqrt(6/(fan_in+fan_out)),dtype=tf.float32))
# self.b=tf.Variable(tf.random.normal([n_filters,],stddev=xavior_std,dtype=tf.float32))
self.b=tf.Variable(tf.zeros([n_filters,],dtype=tf.float32))
self.strides=[1,1,1,1]
def __call__(self,x):
xx=tf.nn.conv2d(x, self.w, strides=self.strides, padding='VALID')+self.b
xx=tf.keras.activations.relu(xx)
y=tf.nn.max_pool2d(xx,ksize=2,strides=2,padding="VALID")
return y
class my_conv_1x1():
def __init__(self,in_chs=32,n_filters=32):
# Determine initialization parameters
k_size=1
fan_in=k_size*k_size*in_chs
fan_out=k_size*k_size*n_filters
#Xavior initialization
self.w=tf.Variable(tf.random.normal([k_size, k_size, in_chs,n_filters],stddev=np.sqrt(2/(fan_in+fan_out)) ,dtype=tf.float32))
# self.w=tf.Variable(tf.random.uniform([k_size, k_size, in_chs,n_filters],minval=-np.sqrt(6/(fan_in+fan_out)),maxval=np.sqrt(6/(fan_in+fan_out)),dtype=tf.float32))
# self.b=tf.Variable(tf.random.normal([n_filters,],stddev=xavior_std,dtype=tf.float32))
self.b=tf.Variable(tf.zeros([n_filters,],dtype=tf.float32))
self.strides=[1,1,1,1]
def __call__(self,x):
y=tf.nn.conv2d(x, self.w, strides=self.strides, padding='VALID')+self.b
# y=tf.keras.activations.relu(y)
return y
class my_dense():
def __init__(self,in_chs,out_chs):
# Determine initialization parameters
fan_in=in_chs
fan_out=out_chs
#Xavior initialization
self.w=tf.Variable(tf.random.normal([in_chs,out_chs],stddev=np.sqrt(2/(fan_in+fan_out)) ,dtype=tf.float32))
# self.w=tf.Variable(tf.random.uniform([in_chs,out_chs],minval=-np.sqrt(6/(fan_in+fan_out)),maxval=np.sqrt(6/(fan_in+fan_out)),dtype=tf.float32))
# self.b=tf.Variable(tf.random.normal([out_chs,],stddev=np.sqrt(2/(fan_in+fan_out)),dtype=tf.float32))
self.b=tf.Variable(tf.zeros([out_chs,],dtype=tf.float32))
def __call__(self,x):
y=tf.matmul(x,self.w)+self.b
return y
########################################## Model ###############################################
class cnn_model():
def __init__(self,way,n_filter=32,drop_rate=0.3):
# super().__init__()
self.cnn1=my_cnn2d(3,1,n_filter)
self.cnn1_1=my_conv_1x1(n_filter,n_filter//2)
self.cnn2=my_cnn2d(3,n_filter//2,n_filter)
self.cnn2_1=my_conv_1x1(n_filter,n_filter//2)
self.cnn3=my_cnn2d(3,n_filter//2,n_filter)
self.cnn3_1=my_conv_1x1(n_filter,n_filter//2)
# self.dense=my_dense(1600,way)
self.dense=my_dense(n_filter//2,way)
self.drop_rate=drop_rate
def __call__(self,x):
xx=self.cnn1(x)
xx=self.cnn1_1(xx)
xx=self.cnn2(xx)
xx=self.cnn2_1(xx)
xx=self.cnn3(xx)
xx=self.cnn3_1(xx)
# global average pooling
xx=tf.nn.avg_pool(xx, ksize=[1]+xx.shape[1:-1]+[1], strides=[1, 1, 1, 1], padding='VALID')
xx=tf.reshape(xx,[xx.shape[0],xx.shape[-1]])
# xx=tf.reshape(xx,[xx.shape[0],-1])
xx=tf.nn.dropout(xx, self.drop_rate)
y=self.dense(xx)
return tf.keras.activations.softmax(y)
def trainable_weights(self):
wt=[]
for ii in [self.cnn1,self.cnn1_1,self.cnn2,self.cnn2_1,self.cnn3,self.cnn3_1,self.dense]:
wt.append(ii.w)
wt.append(ii.b)
return wt
def assign_update(self,tensors):
nn=0
for ww in [self.cnn1,self.cnn1_1,self.cnn2,self.cnn2_1,self.cnn3,self.cnn3_1,self.dense]:
flt_len=np.prod(ww.w.shape)
ww.w.assign_sub(tf.reshape(tensors[nn:nn+flt_len],ww.w.shape))
nn+=flt_len
flt_len=np.prod(ww.b.shape)
ww.b.assign_sub(tf.reshape(tensors[nn:nn+flt_len],ww.b.shape))
nn+=flt_len
def apply_gradients(self,grad):
nn=0
for ww in [self.cnn1,self.cnn1_1,self.cnn2,self.cnn2_1,self.cnn3,self.cnn3_1,self.dense]:
ww.w=ww.w+grad[nn]
ww.b=ww.b+grad[nn+1]
nn+=2
def print_shape(self):
params=0
for idx,ii in enumerate([self.cnn1,self.cnn1_1,self.cnn2,self.cnn2_1,self.cnn3,self.cnn3_1,self.dense]):
params+=np.prod(ii.w.shape)+np.prod(ii.b.shape)
print(f"layer {idx}: w:{ii.w.shape} + b: {ii.b.shape} = {np.prod(ii.w.shape)+np.prod(ii.b.shape)}")
print(f"totla {params} parameters")
class LSTM_model(keras.Model):
def __init__(self,n_units):
super().__init__()
self.initializer = keras.initializers.lecun_uniform()
self.lstm1 = keras.layers.LSTMCell(n_units,kernel_initializer=self.initializer,use_bias=False)
self.lstm2 = keras.layers.LSTMCell(n_units,kernel_initializer=self.initializer,use_bias=False)
self.rnn= keras.layers.RNN([self.lstm1,self.lstm2],return_state=True,stateful=True)
self.dense=keras.layers.Dense(1,kernel_initializer=self.initializer)
def forward(self, x,state=None):
x,*states = self.rnn(x,initial_state=state)
y = self.dense(x)
return y,states