-
Notifications
You must be signed in to change notification settings - Fork 5
/
mobilenet_v1_0.25.py
230 lines (215 loc) · 12.8 KB
/
mobilenet_v1_0.25.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
# mobilenet v1 implemented in RISC-V assembly code
import numpy as np
import keras as kr
from keras.layers import Conv2D, DepthwiseConv2D
from keras.initializers import constant
from machine import machine
# alternatively, uncomment the line below to use the tinyfive package
# from tinyfive.machine import machine
from layers import *
np.random.seed(5) # fix seed for reproducible results
m = machine(mem_size=8000000) # instantiate RISC-V machine with 8MB of memory
# TODO: reduce to 500KB once we use branches to reduce image size
#-------------------------------------------------------------------------------
# run inference (golden reference)
#-------------------------------------------------------------------------------
# mobilenet v1 with 0.25 depth multiplier:
#
# Layer Type Stride In-chan Out-chan In-res Out-res
# ----------------------------------------------------------------
# 1 Conv 3x3 2 3 8 96 48
# 2 Conv DW 3x3 1 8 8 48 48
# 3 Conv 1x1 1 8 16 48 48
# 4 Conv DW 3x3 2 16 16 48 24
# 5 Conv 1x1 1 16 32 24 24
# 6 Conv DW 3x3 1 32 32 24 24
# 7 Conv 1x1 1 32 32 24 24
# 8 Conv DW 3x3 2 32 32 24 12
# 9 Conv 1x1 1 32 64 12 12
# 10 Conv DW 3x3 1 64 64 12 12
# 11 Conv 1x1 1 64 64 12 12
# 12 Conv DW 3x3 2 64 64 12 6
# 13 Conv 1x1 1 64 128 6 6
# 14 Conv DW 3x3 1 128 128 6 6
# 15 Conv 1x1 1 128 128 6 6
# 16 Conv DW 3x3 1 128 128 6 6
# 17 Conv 1x1 1 128 128 6 6
# 18 Conv DW 3x3 1 128 128 6 6
# 19 Conv 1x1 1 128 128 6 6
# 20 Conv DW 3x3 1 128 128 6 6
# 21 Conv 1x1 1 128 128 6 6
# 22 Conv DW 3x3 1 128 128 6 6
# 23 Conv 1x1 1 128 128 6 6
# 24 Conv DW 3x3 2 128 128 6 3
# 25 Conv 1x1 1 128 256 3 3
# 26 Conv DW 3x3 1 256 256 3 3
# 27 Conv 1x1 1 256 256 3 3
# 28 avg pool 1 256 256 3 1
# 29 FC layer 1 256 2 1 1
# generate random input activations and weights
inp = np.random.normal(size=(1, 96, 96, 3)).astype(np.float32)
w1 = np.random.normal(size=(3, 3, 3, 8)).astype(np.float32)
w2 = np.random.normal(size=(3, 3, 8)).astype(np.float32)
w3 = np.random.normal(size=(1, 1, 8, 16)).astype(np.float32)
w4 = np.random.normal(size=(3, 3, 16)).astype(np.float32)
w5 = np.random.normal(size=(1, 1, 16, 32)).astype(np.float32)
w6 = np.random.normal(size=(3, 3, 32)).astype(np.float32)
w7 = np.random.normal(size=(1, 1, 32, 32)).astype(np.float32)
w8 = np.random.normal(size=(3, 3, 32)).astype(np.float32)
w9 = np.random.normal(size=(1, 1, 32, 64)).astype(np.float32)
w10 = np.random.normal(size=(3, 3, 64)).astype(np.float32)
w11 = np.random.normal(size=(1, 1, 64, 64)).astype(np.float32)
w12 = np.random.normal(size=(3, 3, 64)).astype(np.float32)
w13 = np.random.normal(size=(1, 1, 64, 128)).astype(np.float32)
w14 = np.random.normal(size=(3, 3, 128)).astype(np.float32)
w15 = np.random.normal(size=(1, 1, 128, 128)).astype(np.float32)
# skip layers 16-23 because they are the same as layers 14-15; TODO: add layers16-23
w24 = np.random.normal(size=(3, 3, 128)).astype(np.float32)
w25 = np.random.normal(size=(1, 1, 128, 256)).astype(np.float32)
w26 = np.random.normal(size=(3, 3, 256)).astype(np.float32)
w27 = np.random.normal(size=(1, 1, 256, 256)).astype(np.float32)
l1 = Conv2D( 8, 3, strides=2, padding='same', kernel_initializer=constant( w1))(inp)
l2 = DepthwiseConv2D(3, padding='same', depthwise_initializer=constant( w2))( l1)
l3 = Conv2D( 16, 1, kernel_initializer=constant( w3))( l2)
l4 = DepthwiseConv2D(3, strides=2, padding='same', depthwise_initializer=constant( w4))( l3)
l5 = Conv2D( 32, 1, kernel_initializer=constant( w5))( l4)
l6 = DepthwiseConv2D(3, padding='same', depthwise_initializer=constant( w6))( l5)
l7 = Conv2D( 32, 1, kernel_initializer=constant( w7))( l6)
l8 = DepthwiseConv2D(3, strides=2, padding='same', depthwise_initializer=constant( w8))( l7)
l9 = Conv2D( 64, 1, kernel_initializer=constant( w9))( l8)
l10 = DepthwiseConv2D(3, padding='same', depthwise_initializer=constant(w10))( l9)
l11 = Conv2D( 64, 1, kernel_initializer=constant(w11))(l10)
l12 = DepthwiseConv2D(3, strides=2, padding='same', depthwise_initializer=constant(w12))(l11)
l13 = Conv2D( 128, 1, kernel_initializer=constant(w13))(l12)
l14 = DepthwiseConv2D(3, padding='same', depthwise_initializer=constant(w14))(l13)
l15 = Conv2D( 128, 1, kernel_initializer=constant(w15))(l14)
# skip layers 16-23 because they are the same as layers 14-15; TODO: add layers16-23
l24 = DepthwiseConv2D(3, strides=2, padding='same', depthwise_initializer=constant(w24))(l15)
l25 = Conv2D( 256, 1, kernel_initializer=constant(w25))(l24)
l26 = DepthwiseConv2D(3, padding='same', depthwise_initializer=constant(w26))(l25)
l27 = Conv2D( 256, 1, kernel_initializer=constant(w27))(l26)
# TODOs:
# - replace above by keras sequence (or import a reference model from HuggingFace or Keras)
# - use for-loops (for i range(30)) for these 29 layers to clean up the code
# - add the remaining layers 15 to 29
# - add ReLU
# - add biases
#-------------------------------------------------------------------------------
# store activations and weights in memory
#-------------------------------------------------------------------------------
w1_base = 0
w2_base = w1.size * 4
w3_base = w2.size * 4 + w2_base
w4_base = w3.size * 4 + w3_base
w5_base = w4.size * 4 + w4_base
w6_base = w5.size * 4 + w5_base
w7_base = w6.size * 4 + w6_base
w8_base = w7.size * 4 + w7_base
w9_base = w8.size * 4 + w8_base
w10_base = w9.size * 4 + w9_base
w11_base = w10.size * 4 + w10_base
w12_base = w11.size * 4 + w11_base
w13_base = w12.size * 4 + w12_base
w14_base = w13.size * 4 + w13_base
w15_base = w14.size * 4 + w14_base
# skip layers 16-23 because they are the same as layers 14-15; TODO: add layers16-23
w24_base = w15.size * 4 + w15_base
w25_base = w24.size * 4 + w24_base
w26_base = w25.size * 4 + w25_base
w27_base = w26.size * 4 + w26_base
# TODO: optimize the memory footprint of the activations
a1_base = w27.size * 4 + w27_base
a2_base = inp.size * 4 + a1_base
a3_base = l1.numpy().size * 4 + a2_base
a4_base = l2.numpy().size * 4 + a3_base
a5_base = l3.numpy().size * 4 + a4_base
a6_base = l4.numpy().size * 4 + a5_base
a7_base = l5.numpy().size * 4 + a6_base
a8_base = l6.numpy().size * 4 + a7_base
a9_base = l7.numpy().size * 4 + a8_base
a10_base = l8.numpy().size * 4 + a9_base
a11_base = l9.numpy().size * 4 + a10_base
a12_base = l10.numpy().size * 4 + a11_base
a13_base = l11.numpy().size * 4 + a12_base
a14_base = l12.numpy().size * 4 + a13_base
a15_base = l13.numpy().size * 4 + a14_base
# skip layers 16-23 because they are the same as layers 14-15; TODO: add layers16-23
a24_base = l14.numpy().size * 4 + a15_base
a16_base = a24_base # TODO: remove eventually
a25_base = l15.numpy().size * 4 + a24_base
a26_base = l24.numpy().size * 4 + a25_base
a27_base = l25.numpy().size * 4 + a26_base
out_base = l26.numpy().size * 4 + a27_base
a28_base = out_base # TODO: remove eventually
code_start = l27.numpy().size * 4 + out_base
# TODO: clean up, minimize memory size
a16_base = a1_base
a24_base = a1_base
a25_base = a2_base
a26_base = a3_base
a27_base = a4_base
a28_base = a5_base
m.write_f32_vec(np.transpose( w1, axes=[3, 0, 1, 2]).flatten(), w1_base)
m.write_f32_vec(np.transpose( w2, axes=[2, 0, 1]).flatten(), w2_base)
m.write_f32_vec( w3.flatten(), w3_base)
m.write_f32_vec(np.transpose( w4, axes=[2, 0, 1]).flatten(), w4_base)
m.write_f32_vec( w5.flatten(), w5_base)
m.write_f32_vec(np.transpose( w6, axes=[2, 0, 1]).flatten(), w6_base)
m.write_f32_vec( w7.flatten(), w7_base)
m.write_f32_vec(np.transpose( w8, axes=[2, 0, 1]).flatten(), w8_base)
m.write_f32_vec( w9.flatten(), w9_base)
m.write_f32_vec(np.transpose(w10, axes=[2, 0, 1]).flatten(), w10_base)
m.write_f32_vec( w11.flatten(), w11_base)
m.write_f32_vec(np.transpose(w12, axes=[2, 0, 1]).flatten(), w12_base)
m.write_f32_vec( w13.flatten(), w13_base)
m.write_f32_vec(np.transpose(w14, axes=[2, 0, 1]).flatten(), w14_base)
m.write_f32_vec( w15.flatten(), w15_base)
# skip layers 16-23 because they are the same as layers 14-15; TODO: add layers16-23
m.write_f32_vec(np.transpose(w24, axes=[2, 0, 1]).flatten(), w24_base)
m.write_f32_vec( w25.flatten(), w25_base)
m.write_f32_vec(np.transpose(w26, axes=[2, 0, 1]).flatten(), w26_base)
m.write_f32_vec( w27.flatten(), w27_base)
m.write_f32_vec(inp.flatten(), a1_base)
#-------------------------------------------------------------------------------
# run assembly code and check against keras
#-------------------------------------------------------------------------------
def compare_cpu_vs_ref(m, F, R, y_base, ref, trans=False):
"""compare CPU machine versus reference (Keras, PyTorch)"""
if trans==False:
cpu = m.read_f32_vec(y_base, size=R*R*F).reshape(R, R, F)
else:
cpu = np.transpose(m.read_f32_vec(y_base, size=R*R*F).reshape(F, R, R), axes=[1, 2, 0])
m.print_rel_err(cpu, ref.numpy().reshape(R, R, F))
def run_layer(m, ltype, stride, C, F, R, a_base, w_base, y_base, code_start, ref, trans=False):
"""run one layer in assembly and compare resulting tensor with reference 'ref'"""
if ltype=='conv_3x3x3' and stride==2:
conv_3x3x3_stride2(m, F, R, a_base, w_base, y_base)
elif ltype=='dw_conv_3x3' and stride==1:
dw_conv_3x3_stride1(m, C, R, a_base, w_base, y_base, out_chan_first=False)
elif ltype=='dw_conv_3x3' and stride==2:
dw_conv_3x3_stride2(m, C, R, a_base, w_base, y_base, out_chan_first=False)
elif ltype=='conv_1x1' and F<256:
conv_1x1(m, C, F, R, a_base, w_base, y_base, code_start, trans=trans)
else:
conv_1x1_big(m, C, F, R, a_base, w_base, y_base, code_start, S=3, trans=trans)
compare_cpu_vs_ref(m, F, R//stride, y_base, ref, trans)
run_layer(m, 'conv_3x3x3', 2, 3, 8, 96, a1_base, w1_base, a2_base, code_start, l1, trans=True)
run_layer(m, 'dw_conv_3x3', 1, 8, 8, 48, a2_base, w2_base, a3_base, code_start, l2)
run_layer(m, 'conv_1x1', 1, 8, 16, 48, a3_base, w3_base, a4_base, code_start, l3, trans=True)
run_layer(m, 'dw_conv_3x3', 2, 16, 16, 48, a4_base, w4_base, a5_base, code_start, l4)
run_layer(m, 'conv_1x1', 1, 16, 32, 24, a5_base, w5_base, a6_base, code_start, l5, trans=True)
run_layer(m, 'dw_conv_3x3', 1, 32, 32, 24, a6_base, w6_base, a7_base, code_start, l6)
run_layer(m, 'conv_1x1', 1, 32, 32, 24, a7_base, w7_base, a8_base, code_start, l7, trans=True)
run_layer(m, 'dw_conv_3x3', 2, 32, 32, 24, a8_base, w8_base, a9_base, code_start, l8)
run_layer(m, 'conv_1x1', 1, 32, 64, 12, a9_base, w9_base, a10_base, code_start, l9, trans=True)
run_layer(m, 'dw_conv_3x3', 1, 64, 64, 12, a10_base, w10_base, a11_base, code_start, l10)
run_layer(m, 'conv_1x1', 1, 64, 64, 12, a11_base, w11_base, a12_base, code_start, l11, trans=True)
run_layer(m, 'dw_conv_3x3', 2, 64, 64, 12, a12_base, w12_base, a13_base, code_start, l12)
run_layer(m, 'conv_1x1', 1, 64, 128, 6, a13_base, w13_base, a14_base, code_start, l13, trans=True)
run_layer(m, 'dw_conv_3x3', 1, 128, 128, 6, a14_base, w14_base, a15_base, code_start, l14)
run_layer(m, 'conv_1x1', 1, 128, 128, 6, a15_base, w15_base, a16_base, code_start, l15, trans=True)
# skip layers 16-23 because they are identical to layers 14-15; TODO: add them
run_layer(m, 'dw_conv_3x3', 2, 128, 128, 6, a24_base, w24_base, a25_base, code_start, l24)
run_layer(m, 'conv_1x1', 1, 128, 256, 3, a25_base, w25_base, a26_base, code_start, l25, trans=True)
run_layer(m, 'dw_conv_3x3', 1, 256, 256, 3, a26_base, w26_base, a27_base, code_start, l26)
run_layer(m, 'conv_1x1', 1, 256, 256, 3, a27_base, w27_base, a28_base, code_start, l27)