-
Notifications
You must be signed in to change notification settings - Fork 1
/
cifar_keras_inference.py
354 lines (274 loc) · 13.2 KB
/
cifar_keras_inference.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
'''
Peforms 8-bit INT inference of a simple Keras Adapted CNN on the CIFAR-10 dataset.
Gets to 82.46% test accuracy after 89 epochs using tensorflow backend
Example:
Inference Mode:
python cifar_keras.py -w weights_int_8bit_signed.hdf5
#load quantized weights
Inference Mode and print intermediate layer pkl files
python cifar_keras.py -w weights_int_8bit_signed.hdf5 -p 1
'''
from __future__ import print_function
import numpy as np
np.random.seed(0) # for reproducibility
import tensorflow as tf
import keras.backend as K
import csv
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Lambda, Dense, Dropout, Activation, BatchNormalization, MaxPooling2D, Conv2D
from keras.layers import Flatten
from keras.optimizers import SGD, Adam, RMSprop
from keras.callbacks import LearningRateScheduler, ModelCheckpoint
from keras.utils import np_utils
from keras.models import Model
from keras.regularizers import l2
from binary_ops import relu_layer as relu_layer_op
from binary_ops import softmax_layer as softmax_layer_op
from binary_ops import floor_func as floor_func_op
from binary_layers import BinaryDense, BinaryConv2D
from matplotlib import pyplot as plt
import argparse
import math
from DAC import dac_param, give_an_input_get_analog_output_dac
from ADC import adc_param, give_vmav_get_yout
from MAV import mav_transfer, give_weight_get_vmav
from cifar10_inference import cim_conv, conv
weight_hdf5 = 'quantized_0720.hdf5'
def relu_layer(x):
return relu_layer_op(x)
def softmax_layer(x):
return softmax_layer_op(x)
def floor_func(x,divisor):
return floor_func_op(x,divisor)
def clip_func(x):
low_values_flags = x < -127
x[low_values_flags] = 0
high_values_flags = x > 127
x[high_values_flags] = 128
return x
batch_size = 32
epochs = 100
channels = 3
img_rows = 32
img_cols = 32
classes = 10
use_bias = False
# Batch Normalization
epsilon = 1e-6
momentum = 0.9
weight_decay = 0.0004
# construct the argument parse and parse the arguments
ap = argparse.ArgumentParser()
ap.add_argument("-w", "--weights", type=str,
help="(optional) path to weights file")
ap.add_argument("-p", "--print_layers", type=int, default=-1,
help="(optional) To print intermediate layer pkl files")
args = vars(ap.parse_args())
######################
#### DATA ####
######################
from data.cifar_test_yarray import y_test
from data.cifar_test_xarray import X_test
Y_test = np_utils.to_categorical(y_test, classes)
# padding X_test with zeros
X_test_padded = np.zeros(shape=(X_test.shape[0], X_test.shape[1], X_test.shape[2]+2, X_test.shape[3]+2))
X_test_padded[:X_test.shape[0], :X_test.shape[1], 1:X_test.shape[2]+1, 1:X_test.shape[3]+1] = X_test
# print(X_test_padded[0])
# print(X_test_padded[:1].shape, 'test samples')
# print(Y_test[:1].shape, 'test samples values')
X_test_small = X_test_padded[:1000]
Y_test_small = Y_test[:1000]
# Weight
weight_data_conv1 = []
with open('converted_weight_CONV1_dec.txt', 'r') as weight_file:
reader = csv.reader(weight_file)
for row in reader:
row.pop()
weight_data_conv1.append( [int(i) for i in row] )
weight_file.close()
weight_data_conv2 = []
with open('converted_weight_CONV2_dec.txt', 'r') as weight_file:
reader = csv.reader(weight_file)
for row in reader:
row.pop()
weight_data_conv2.append( [int(i) for i in row] )
weight_file.close()
# for i in range(0, len(weight_data)):
# weight_data[i] = int(weight_data[0][i])
# print(weight_data[0][0])
# x_test = np.random.randint(0,256,size=(3,32,32))
# weight_data1 = np.random.randint(-126,126,size=(2,27))
# print(weight_data1)
def network(X_test_data, Y_test_data):
layers_array = ["conv1"]
with open('max_dict.csv', mode='r') as infile:
reader = csv.reader(infile, delimiter=',')
data_read = [row for row in reader]
conv_scale = []
for i in range(0, 4):
conv_scale.append(math.floor(127 / float(data_read[i * 2][1])))
print(conv_scale[i])
accr_list = []
top_5_acc = []
model = Sequential()
# Conv1, Scaling1 and ReLU1
# model.add(Conv2D(32, kernel_size=(3, 3), input_shape=(channels, img_rows, img_cols), data_format='channels_first',
# kernel_initializer='he_normal', padding='same', use_bias=False, name='conv1'))
model.add(Lambda(lambda x: floor_func(x, conv_scale[0]),
name='scaling1', input_shape=(32,32,32))) ## Dividing by 27 (MAV) and 18.296 (Instead of 128), so need to multiply by factor of 7 in gain stage
model.add(Activation(relu_layer, name='act_conv1'))
# Conv2, Scaling2 and ReLU2
model.add(
Conv2D(32, kernel_size=(3, 3), data_format='channels_first', kernel_initializer='he_normal', padding='same',
use_bias=use_bias, name='conv2'))
model.add(Lambda(lambda x: floor_func(x, conv_scale[1]),
name='scaling2')) ## Dividing by 288 (MAV) and 1 (Instead of 128), so need to multiply by factor of 128 in gain stage
model.add(Activation(relu_layer, name='act_conv2'))
# Pool1
model.add(MaxPooling2D(pool_size=(2, 2), name='pool1', data_format='channels_first'))
# Conv3, Scaling3 and ReLU3
model.add(
Conv2D(64, kernel_size=(3, 3), data_format='channels_first', kernel_initializer='he_normal', padding='same',
use_bias=use_bias, name='conv3'))
model.add(Lambda(lambda x: floor_func(x, conv_scale[2]),
name='scaling3')) ## Dividing by 288 (MAV) and 2 (Instead of 128), so need to multiply by factor of 64 in gain stage
model.add(Activation(relu_layer, name='act_conv3'))
# Conv4, Scaling4 and ReLU4
model.add(
Conv2D(64, kernel_size=(3, 3), data_format='channels_first', kernel_initializer='he_normal', padding='same',
use_bias=use_bias, name='conv4'))
model.add(Lambda(lambda x: floor_func(x, conv_scale[3]),
name='scaling4')) ## Dividing by 576 (MAV) and 1 (Instead of 128), so need to multiply by factor of 128 in gain stage
model.add(Activation(relu_layer, name='act_conv4'))
# Pool2
model.add(MaxPooling2D(pool_size=(2, 2), name='pool2', data_format='channels_first'))
model.add(Flatten())
# model.add(Lambda(lambda x: x*6, name='scaling_fc'))
# FC1, Batch Normalization and ReLU5
model.add(Dense(512, use_bias=True, name='FC1', kernel_initializer='he_normal'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn1'))
model.add(Activation(relu_layer, name='act_fc1'))
# FC2, Batch Normalization and ReLU6
model.add(Dense(classes, use_bias=True, name='FC2', kernel_initializer='he_normal'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn2'))
model.add(Activation(softmax_layer, name='act_fc2'))
# Optimizers
opt = RMSprop(lr=0.0001, decay=1e-6)
model.compile(loss='squared_hinge', optimizer=opt, metrics=['accuracy', 'top_k_categorical_accuracy'])
# model.compile('adam', 'categorical_crossentropy', ['accuracy', 'top_k_categorical_accuracy'])
model.build()
model.summary()
model.load_weights(weight_hdf5, by_name=True)
score = model.evaluate(X_test_data, Y_test_data, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
print('top-k accuracy:', score[2])
accr_list.append(score[1])
top_5_acc.append(score[2])
## LAYER OUTPUTS TO DUMP
if args["print_layers"] > 0:
for i in layers_array:
intermediate_layer_model = Model(inputs=model.input, outputs=model.get_layer(i).output)
intermediate_output = intermediate_layer_model.predict([X_test_data])
file_name = "output/" + i + ".pkl"
print("Dumping layer {} outputs to file {}".format(i, file_name))
intermediate_output.dump(file_name)
def network_f(X_test_data, Y_test_data):
layers_array = ["conv1"]
with open('max_dict.csv', mode='r') as infile:
reader = csv.reader(infile, delimiter=',')
data_read = [row for row in reader]
conv_scale = []
for i in range(0, 4):
conv_scale.append(math.floor(127 / float(data_read[i * 2][1])))
print(conv_scale[i])
accr_list = []
top_5_acc = []
model = Sequential()
# Conv1, Scaling1 and ReLU1
model.add(Conv2D(32, kernel_size=(3, 3), input_shape=(channels, img_rows, img_cols), data_format='channels_first',
kernel_initializer='he_normal', padding='same', use_bias=False, name='conv1'))
model.add(Lambda(lambda x: floor_func(x, conv_scale[0]),
name='scaling1')) ## Dividing by 27 (MAV) and 18.296 (Instead of 128), so need to multiply by factor of 7 in gain stage
model.add(Activation(relu_layer, name='act_conv1'))
# Conv2, Scaling2 and ReLU2
model.add(
Conv2D(32, kernel_size=(3, 3), data_format='channels_first', kernel_initializer='he_normal', padding='same',
use_bias=use_bias, name='conv2'))
model.add(Lambda(lambda x: floor_func(x, conv_scale[1]),
name='scaling2')) ## Dividing by 288 (MAV) and 1 (Instead of 128), so need to multiply by factor of 128 in gain stage
model.add(Activation(relu_layer, name='act_conv2'))
# Pool1
model.add(MaxPooling2D(pool_size=(2, 2), name='pool1', data_format='channels_first'))
# Conv3, Scaling3 and ReLU3
model.add(
Conv2D(64, kernel_size=(3, 3), data_format='channels_first', kernel_initializer='he_normal', padding='same',
use_bias=use_bias, name='conv3'))
model.add(Lambda(lambda x: floor_func(x, conv_scale[2]),
name='scaling3')) ## Dividing by 288 (MAV) and 2 (Instead of 128), so need to multiply by factor of 64 in gain stage
model.add(Activation(relu_layer, name='act_conv3'))
# Conv4, Scaling4 and ReLU4
model.add(
Conv2D(64, kernel_size=(3, 3), data_format='channels_first', kernel_initializer='he_normal', padding='same',
use_bias=use_bias, name='conv4'))
model.add(Lambda(lambda x: floor_func(x, conv_scale[3]),
name='scaling4')) ## Dividing by 576 (MAV) and 1 (Instead of 128), so need to multiply by factor of 128 in gain stage
model.add(Activation(relu_layer, name='act_conv4'))
# Pool2
model.add(MaxPooling2D(pool_size=(2, 2), name='pool2', data_format='channels_first'))
model.add(Flatten())
# model.add(Lambda(lambda x: x*6, name='scaling_fc'))
# FC1, Batch Normalization and ReLU5
model.add(Dense(512, use_bias=True, name='FC1', kernel_initializer='he_normal'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn1'))
model.add(Activation(relu_layer, name='act_fc1'))
# FC2, Batch Normalization and ReLU6
model.add(Dense(classes, use_bias=True, name='FC2', kernel_initializer='he_normal'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn2'))
model.add(Activation(softmax_layer, name='act_fc2'))
# Optimizers
opt = RMSprop(lr=0.0001, decay=1e-6)
model.compile(loss='squared_hinge', optimizer=opt, metrics=['accuracy', 'top_k_categorical_accuracy'])
# model.compile('adam', 'categorical_crossentropy', ['accuracy', 'top_k_categorical_accuracy'])
model.build()
model.summary()
model.load_weights(weight_hdf5, by_name=True)
score = model.evaluate(X_test_data, Y_test_data, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
print('top-k accuracy:', score[2])
accr_list.append(score[1])
top_5_acc.append(score[2])
## LAYER OUTPUTS TO DUMP
if args["print_layers"] > 0:
for i in layers_array:
intermediate_layer_model = Model(inputs=model.input, outputs=model.get_layer(i).output)
intermediate_output = intermediate_layer_model.predict([X_test_data])
file_name = "output/" + i + ".pkl"
print("Dumping layer {} outputs to file {}".format(i, file_name))
intermediate_output.dump(file_name)
# testing
def __main__():
# print('nothing here')
next_layer_input = conv(X_test_small, weight_data_conv1, 270)
# print(next_layer_input)
# print('shape = ' + str(next_layer_input.shape))
# print('max = ' + str(np.amax(next_layer_input)))
# padding for conv2
# next_layer_input_padding = np.zeros(shape=(
# next_layer_input.shape[0], next_layer_input.shape[1], next_layer_input.shape[2] + 2, next_layer_input.shape[3] + 2))
# next_layer_input_padding[:next_layer_input.shape[0], :next_layer_input.shape[1], 1:next_layer_input.shape[2] + 1,
# 1:next_layer_input.shape[3] + 1] = next_layer_input
# next_layer_input_con2 = conv(next_layer_input_padding, weight_data_conv2, 128)
# print('conv2 shape = ' + str(next_layer_input_con2.shape))
# print('conv2 max = ' + str(np.amax(next_layer_input_con2)))
# print(next_layer_input)
# print('shape = ' + str(next_layer_input.shape)) np.zeros(shape=(1,32,32,32))
network(X_test_data=next_layer_input, Y_test_data=Y_test_small)
# print(X_test_small.shape, 'test samples', Y_test_small.shape, 'test_sample_value')
# print('shape = ' + str(next_layer_input.shape))
# network_f(X_test_data=X_test[:1], Y_test_data=Y_test[:1])
if __name__ == "__main__":
# Actually run your code in here
__main__()