-
Notifications
You must be signed in to change notification settings - Fork 1
/
layers.py
412 lines (330 loc) · 14.4 KB
/
layers.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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
from inputs import Input
class Layer(object):
"""Base layers class. This is the class from whicn all layers inherit.
A layer is class implementing common neural network operations, e.g.
convolutional layer, recurrent layer, etc.
Args:
parameters: list, containing all parameters e.g. weights, biases
input_tensor: like-tensor, input tensor
input_shape: The shape of the tensor
"""
def __init__(self):
self.params = []
self.input_tensor = None
self.input_shape = None
self.output_shape = None
def add_params(self, params):
"""Input params that layer need to add, e.g. weights, biases
Args:
parameters: list, layer params
"""
params.set_trainable(trainable=True)
self.params.append(params)
def get_params(self):
"""get all params that the layer need to update
"""
return self.params
def set_input_shape(self, input_shape):
"""setup the input tensor shape for layer
Args:
input_shape: int list, input tensor shape
"""
self.input_shape = input_shape
def get_input_shape(self):
"""get input tensor shape
"""
return self.input_shape[0]
def set_output_shape(self, output_shape):
"""setup output tensor shape for layers
Args:
output_shape: int list, output shape
"""
self.output_shape = output_shape
def get_output_shape(self):
"""get output tensor shape, it will be realized from sub-class
"""
raise NotImplementedError
def init_params(self):
"""initialize layer parameters
"""
# raise NotImplementedError
pass
def __call__(self, input_pl):
"""callable object, a operator for model
Args:
input_pl: input placeholder
"""
if not isinstance(input_pl, Input):
raise ValueError('Input placeholder mus be same type with Input')
# setup input_placeholder as current layer input, get shape
self.set_input_shape(input_pl.shape)
# get output_shape = input_pl_shape, takes input_pl_shape as output shape of a layer
output_shape = self.get_output_shape()
# create a output_placeholder
return Input(output_shape, [input_pl], self)
class Add(Layer):
"""Layer that adds a list of inputs. It takes as input a list of tensors, all of the same shape, and returns
a single tensor (also of the same shape).
"""
def __init__(self):
super(Add, self).__init__()
def get_output_shape(self):
"""inherited class from Layer to compute output shape
"""
# input shape of a layer is equal to output shape
output_shape = self.get_input_shape()
# setup output shape
self.set_output_shape(output_shape)
return output_shape
def forward(self, input_tensor1, input_tensor2):
return input_tensor1.__add__(input_tensor2)
def __call__(self, input_pl1, input_pl2):
"""callable method to add 2 layer
"""
if (not isinstance(input_pl1, Input)) and (not isinstance(input_pl2, Input)):
raise ValueError('Layer1 and layer2 must be same type')
if input_pl1.shape != input_pl2.shape:
raise ValueError('Layer1 and layer2 must be same shape')
# input_pl is current layer input
self.set_input_shape(input_pl1.shape)
output_shape = self.get_output_shape()
# create output placeholder
return Input(output_shape, [input_pl1, input_pl2], self)
class Flatten(Layer):
"""Flattens the input. Does not affect the batch size
"""
def __init__(self):
super(Flatten, self).__init__()
def get_output_shape(self):
input_shape = self.get_input_shape()
# output_shape = (batch_size, channel*height*width)
output_shape = (input_shape[0], np.prod(input_shape[1:]))
self.set_output_shape(output_shape)
return output_shape
def forword(self, input_tensor):
return input_tensor.flatten()
class Reshape(Layer):
"""Reshapes an output to a certain shape.
Args:
target_shape: integers, Target shape.
"""
def __init__(self, target_shape):
self.target_shape = target_shape
super(Reshape, self).__init__()
def get_output_shape(self):
output_shape = self.target_shape
self.set_output_shape(output_shape)
return output_shape
def forward(self, input_tensor):
return input_tensor.reshape(target_shape)
class Activation(Layer):
"""Applies an activation function to an output.
Activation subclass inherts from base class Layer
Args:
activation: Activation function, string name of built-in activation function, such as "relu".
"""
def __init__(self, activation):
# In Python 3, we could use super()__init__() syntaxe to inherts baseclass
super(Activation, self).__init__()
self.activation = activation
def get_output_shape(self):
"""This method is inherited from base-class
"""
output_shape = self.get_input_shape()
self.set_output_shape(output_shape)
return output_shape
def forward(self, input_tensor):
"""forward propagation function
Args:
input_tensor: Tensor, input tensor fo shape [batch_size, c, h, w]
"""
if self.activation == 'relu':
return input_tensor.relu()
elif self.activation == 'sigmoid':
return input_tensor.sigmoid()
elif self.activation == 'tanh':
return input_tensor.tanh()
elif self.activation == 'softmax':
return input_tensor.softmax()
else:
raise ValueError('There are not %s activation function, default activations are relu, sigmoid, tanh and softmax')
class Dropout(Layer):
"""Applies Dropout to the input.
Args:
rate: Float between 0 and 1. Fraction of the input units to drop.
"""
def __init__(self, rate):
super(Dropout, self).__init__()
self.rate = rate
def get_output_shape():
output_shape = self.get_input_shape()
self.set_output_shape(output_shape)
return output_shape
def forward(self, input_tensor):
return input_tensor.dropout(self.rate)
class Conv2D(Layer):
"""2D convolution layer
Args:
filter_nums: Integer, the number of filters in the convolution
ksize: Tuple or list of integer, the length of the convolution window.
stride: Integer, the stride length of the convolution.
pad: Integer, the padding size
input_shape: Tuple of list of integer, the input tensor shape, default value is None
weight_initializer: Initializer for the `kernel` weights matrix.
bias_initializer: Initializer for the bias vector.
"""
def __init__(self, filter_nums, ksize, stride=1, pad=0, input_shape=None, weight_initializer='normal', bias_initializer='zeros'):
# inherite base class from Layer
super(Conv2D, self).__init__()
self.filter_nums = filter_nums
self.ksize = ksize
self.stride =stride
self.pad = pad
if input_shape is not None:
input_c, input_h, input_w = input_shape
self.set_input_shape((None, input_c, input_h, input_w))
# self.weight_vals = initializer.get(weight_initializer)
# self.bias_vals = initializer.get(weight_initializer)
self.weight = None
self.bias = None
def get_output_shape(self):
"""Should to compute output conv layer shape
"""
input_nums, input_c, input_h, input_w = self.get_input_shape()
# comput output_h and output_w from function get_conv_output_shape
output_h, output_w = get_conv_output_shape(input_h, input_w, self.ksize, self.stride, self.pad)
output_shape = (input_nums, self.filter_nums, output_h, output_w)
self.set_output_shape(output_shape)
return output_shape
def init_params(self):
"""sub-class method that inherit from base class Layer
"""
kernel_h, kernel_w = self.ksize
input_c = self.get_input_shape()[1]
# weight_vals = initializer.get(weight_initializer)
# bias_vals = initializer.get(bias_initializer)
weight_vals = np.random.randn(input_c * kernel_h * kernel_w, self.filter_nums) * np.sqrt(2.0 / input_c * kernel_h * kernel_w)
bias_vals = np.zeros(self.filter_nums)
self.weight = Tensor(weight_vals, auto_grad=True)
self.bias = Tensor(bias_vals, auto_grad=True)
# apend weight and bias into self.params
self.params.append(self.weight)
self.params.append(self.bias)
def forward(self, input_tensor):
"""forward propagation
Args:
input_tensor: Tensor with shape of [batch_size, c, h, w]
Returns:
output tensor of shape [batch_size, c, output_h, output_w]
"""
input_nums, input_c, input_h, input_w = input_tensor.shape
output_h, output_w = get_conv_output_shape(input_h, input_w, self.ksize, self.stride, self.pad)
# determine weight and bias are existed
if (self.weight is None) or (self.bias is None):
self.get_input_shape(input_nums, input_c, input_h, input_w)
self.get_output_shape()
self.init_params()
# Expand input data into a two-dimensional array of shape [input_nums*output_h*output_w, input_c*kernel_h*kernel_w]
matrix = input_tensor.tensor_to_matrix(self.ksize, self.stride, self.pad)
# shape [input_num*output_h*output_h, filter_nums]
output = matrix.dot(self.weight)
output = output + self.bias.expand(0, matrix.data.shape[0]) # [input_nums * output_h * output_w, filter_nums]
return output.reshape(input_nums, output_h, ouput_w, -1).transpose(0, 3, 1, 2) # [input_nums, filter_nums, output_h, output_w]
class MaxPooling(Layer):
"""Max pooling operation for spatial data.
Attributs:
pool_size: Integer, size of the max pooling windows.
strides: Integer, Factor by which to downscale.
pad: Integer, the padding size
input_shape: shape tupel or list
"""
def __init__(self, pool_size, stride=1, pad=0, input_shape=None):
super(MaxPooling, self).__init__()
self.pool_size = pool_size
self.stride = stride
self.pad = pad
self.input_shape = input_shape
if input_shape is not None:
input_c, input_h, input_w = input_shape
self.set_input_shape((None, input_c, input_h, input_w))
def get_output_shape(self):
"""Should to compute output maxpooling layer shape
"""
input_nums, input_c, input_h, input_w = self.get_input_shape()
# comput output_h and output_w from function get_conv_output_shape
output_h, output_w = get_conv_output_shape(input_h, input_w, self.pool_size, self.stride, self.pad)
output_shape = (input_nums, self.filter_nums, output_h, output_w)
self.set_output_shape(output_shape)
return output_shape
def forward(self, input_tensor):
"""forward propagation
Args:
input_tensor: Tensor with shape of [batch_size, c, h, w]
"""
# setup input_tensor shape
self.set_input_shape(input_tensor.data.shape)
output_shape = self.get_output_shape()
input_nums, input_c, input_h, input_w = input_tensor.data.shape
# compute output h and w from the function get_conv_output_shape
output_h, output_w = get_conv_output_shape(input_h, input_w, self.pool_size, self.stride, self.pad)
# Expand input data into a two-dimensional array of shape [input_nums*output_h*output_w, input_c*kernel_h*kernel_w]
matrix = input_tensor.tensor_to_matrix(self.pool_size, self.stride, self.pad)
matrix = matrix.reshape((-1, self.pool_size[0] * self.pool_size[1]))
output = matrix.max()
return output.reshape((input_nums, output_h, output_w, input_c)).transpose((0, 3, 1, 2))
class Dense(Layer):
"""Regular densely-connected NN layer.
Dense implements the operation: output = dot(inputs, weight) + bias
Args:
output_dim: Integer, dim of output = input_shape[-1]
input_shape: shape tupel or list
weight_initializer: String
bias_initializer: String
"""
def __init__(self, output_dim, input_shape=None, weight_initializer='normal', bias_initializer='zeros'):
super(Dense, self).__init__()
self.output_dim = output_dim
self.weight = None
self.bias = None
if input_shape is not None:
self.set_input_shape((None,) + input_shape)
def get_output_shape(self):
input_shape = self.get_input_shape()
output_shape = input_shape[:-1] + (self.output_dim, )
self.set_output_shape(output_shape)
return output_shape
def init_params(self):
# get input and output dim
input_dim = self.get_input_shape()[-1]
output_dim = self.get_output_shape()[-1]
# init weight and bias
# weight_vals = initializer.get(weight_initializer)
# bais_vals = initializer(bias_initializer)
weight_vals = np.random.random(input_dim, output_dim) * np.sqrt(2.0 / input_dim)
bias_vals = np.zeros((output_dim))
self.weight = Tensor(weight_vals, auto_grad=True)
self.bais = Tensor(bias_vals, auto_grad=True)
# add params into param_list
self.add_params(weight)
self.add_params(bias)
def forward(self, input_tensor):
"""forward compute fot sequence and graph model
"""
if self.weight is None:
self.set_input_shape(input_tensor.data.shape)
output_shape = self.get_output_shape()
self.init_params()
if len(input_tensor.data.shape) == 2:
batch_size, _ = input_tensor.data.shape
# kernel = dot(inputs, weights)
output = input_tensor.data.dot(self.weight) + self.bias.expand(0, batch_size)
return output
elif len(input_tensor.data.shape) == 3:
batch_size, repeats, _ = input_tensor.data.shape
kernel = input_tensor.data.dot(self.weight)
output = kernel + self.bias.expand(0, repeats).expand(0, batch_size)
return output