-
Notifications
You must be signed in to change notification settings - Fork 12
/
Copy pathconvolution.py
208 lines (167 loc) · 8.1 KB
/
convolution.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
import numpy as np
import pickle
from os import path, makedirs, remove
from utilities.utils import pad_inputs
from utilities.initializers import glorot_uniform
from utilities.settings import get_layer_num, inc_layer_num
class Convolution:
def __init__(self, filters, kernel_shape=(3, 3), padding='valid', stride=1, name=None):
self.params = {
'filters': filters,
'padding': padding,
'kernel_shape': kernel_shape,
'stride': stride
}
self.cache = {}
self.rmsprop_cache = {}
self.momentum_cache = {}
self.grads = {}
self.has_units = True
self.name = name
self.type = 'conv'
def has_weights(self):
return self.has_units
def save_weights(self, dump_path):
dump_cache = {
'cache': self.cache,
'grads': self.grads,
'momentum': self.momentum_cache,
'rmsprop': self.rmsprop_cache
}
save_path = path.join(dump_path, self.name+'.pickle')
makedirs(path.dirname(save_path), exist_ok=True)
remove(save_path)
with open(save_path, 'wb') as d:
pickle.dump(dump_cache, d)
def load_weights(self, dump_path):
if self.name is None:
self.name = '{}_{}'.format(self.type, get_layer_num(self.type))
inc_layer_num(self.type)
read_path = path.join(dump_path, self.name+'.pickle')
with open(read_path, 'rb') as r:
dump_cache = pickle.load(r)
self.cache = dump_cache['cache']
self.grads = dump_cache['grads']
self.momentum_cache = dump_cache['momentum']
self.rmsprop_cache = dump_cache['rmsprop']
def conv_single_step(self, input, W, b):
'''
Function to apply one filter to input slice.
:param input:[numpy array]: slice of input data of shape (f, f, n_C_prev)
:param W:[numpy array]: One filter of shape (f, f, n_C_prev)
:param b:[numpy array]: Bias value for the filter. Shape (1, 1, 1)
:return:
'''
return np.sum(np.multiply(input, W)) + float(b)
def forward_propagate(self, X, save_cache=False):
'''
:param X:
:param save_cache:
:return:
'''
if self.name is None:
self.name = '{}_{}'.format(self.type, get_layer_num(self.type))
inc_layer_num(self.type)
(num_data_points, prev_height, prev_width, prev_channels) = X.shape
filter_shape_h, filter_shape_w = self.params['kernel_shape']
if 'W' not in self.params:
shape = (filter_shape_h, filter_shape_w, prev_channels, self.params['filters'])
self.params['W'], self.params['b'] = glorot_uniform(shape=shape)
if self.params['padding'] == 'same':
pad_h = int(((prev_height - 1)*self.params['stride'] + filter_shape_h - prev_height) / 2)
pad_w = int(((prev_width - 1)*self.params['stride'] + filter_shape_w - prev_width) / 2)
n_H = prev_height
n_W = prev_width
else:
pad_h = 0
pad_w = 0
n_H = int((prev_height - filter_shape_h) / self.params['stride']) + 1
n_W = int((prev_width - filter_shape_w) / self.params['stride']) + 1
self.params['pad_h'], self.params['pad_w'] = pad_h, pad_w
Z = np.zeros(shape=(num_data_points, n_H, n_W, self.params['filters']))
X_pad = pad_inputs(X, (pad_h, pad_w))
for i in range(num_data_points):
x = X_pad[i]
for h in range(n_H):
for w in range(n_W):
vert_start = self.params['stride'] * h
vert_end = vert_start + filter_shape_h
horiz_start = self.params['stride'] * w
horiz_end = horiz_start + filter_shape_w
for c in range(self.params['filters']):
x_slice = x[vert_start: vert_end, horiz_start: horiz_end, :]
Z[i, h, w, c] = self.conv_single_step(x_slice, self.params['W'][:, :, :, c],
self.params['b'][:, :, :, c])
if save_cache:
self.cache['A'] = X
return Z
def back_propagate(self, dZ):
'''
:param dZ:
:return:
'''
A = self.cache['A']
filter_shape_h, filter_shape_w = self.params['kernel_shape']
pad_h, pad_w = self.params['pad_h'], self.params['pad_w']
(num_data_points, prev_height, prev_width, prev_channels) = A.shape
dA = np.zeros((num_data_points, prev_height, prev_width, prev_channels))
self.grads = self.init_cache()
A_pad = pad_inputs(A, (pad_h, pad_w))
dA_pad = pad_inputs(dA, (pad_h, pad_w))
for i in range(num_data_points):
a_pad = A_pad[i]
da_pad = dA_pad[i]
for h in range(prev_height):
for w in range(prev_width):
vert_start = self.params['stride'] * h
vert_end = vert_start + filter_shape_h
horiz_start = self.params['stride'] * w
horiz_end = horiz_start + filter_shape_w
for c in range(self.params['filters']):
a_slice = a_pad[vert_start: vert_end, horiz_start: horiz_end, :]
da_pad[vert_start:vert_end, horiz_start:horiz_end, :] += self.params['W'][:, :, :, c] * dZ[i, h, w, c]
self.grads['dW'][:, :, :, c] += a_slice * dZ[i, h, w, c]
self.grads['db'][:, :, :, c] += dZ[i, h, w, c]
dA[i, :, :, :] = da_pad[pad_h: -pad_h, pad_w: -pad_w, :]
return dA
def init_cache(self):
cache = dict()
cache['dW'] = np.zeros_like(self.params['W'])
cache['db'] = np.zeros_like(self.params['b'])
return cache
def momentum(self, beta=0.9):
if not self.momentum_cache:
self.momentum_cache = self.init_cache()
self.momentum_cache['dW'] = beta * self.momentum_cache['dW'] + (1 - beta) * self.grads['dW']
self.momentum_cache['db'] = beta * self.momentum_cache['db'] + (1 - beta) * self.grads['db']
def rmsprop(self, beta=0.999, amsprop=True):
if not self.rmsprop_cache:
self.rmsprop_cache = self.init_cache()
new_dW = beta * self.rmsprop_cache['dW'] + (1 - beta) * (self.grads['dW']**2)
new_db = beta * self.rmsprop_cache['db'] + (1 - beta) * (self.grads['db']**2)
if amsprop:
self.rmsprop_cache['dW'] = np.maximum(self.rmsprop_cache['dW'], new_dW)
self.rmsprop_cache['db'] = np.maximum(self.rmsprop_cache['db'], new_db)
else:
self.rmsprop_cache['dW'] = new_dW
self.rmsprop_cache['db'] = new_db
def apply_grads(self, learning_rate=0.001, l2_penalty=1e-4, optimization='adam', epsilon=1e-8,
correct_bias=False, beta1=0.9, beta2=0.999, iter=999):
if optimization != 'adam':
self.params['W'] -= learning_rate * (self.grads['dW'] + l2_penalty * self.params['W'])
self.params['b'] -= learning_rate * (self.grads['db'] + l2_penalty * self.params['b'])
else:
if correct_bias:
W_first_moment = self.momentum_cache['dW'] / (1 - beta1 ** iter)
b_first_moment = self.momentum_cache['db'] / (1 - beta1 ** iter)
W_second_moment = self.rmsprop_cache['dW'] / (1 - beta2 ** iter)
b_second_moment = self.rmsprop_cache['db'] / (1 - beta2 ** iter)
else:
W_first_moment = self.momentum_cache['dW']
b_first_moment = self.momentum_cache['db']
W_second_moment = self.rmsprop_cache['dW']
b_second_moment = self.rmsprop_cache['db']
W_learning_rate = learning_rate / (np.sqrt(W_second_moment) + epsilon)
b_learning_rate = learning_rate / (np.sqrt(b_second_moment) + epsilon)
self.params['W'] -= W_learning_rate * (W_first_moment + l2_penalty * self.params['W'])
self.params['b'] -= b_learning_rate * (b_first_moment + l2_penalty * self.params['b'])