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
|
#
# Copyright (C) 2018 The Android Open Source Project
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
layout = BoolScalar("layout", False) # NHWC
# TEST 1: TRANSPOSE_CONV2D, pad = valid, stride = 2
i1 = Input("op1", "TENSOR_FLOAT32", "{1, 2, 2, 1}") # input 0
w1 = Parameter("op2", "TENSOR_FLOAT32", "{2, 3, 3, 1}", [1, 3, 5, 7, 9, 11, 13, 15, 17, 2, 4, 6, 8, 10, 12, 14, 16, 18]) # weight
b1 = Parameter("op3", "TENSOR_FLOAT32", "{2}", [-1.5, -2]) # bias
s1 = Int32Vector("shape", [1, 5, 5, 2]) # output shape
act = Int32Scalar("act", 0) # act = none
o1 = Output("op4", "TENSOR_FLOAT32", "{1, 5, 5, 2}") # output
Model().Operation("TRANSPOSE_CONV_2D", i1, w1, b1, s1, 2, 2, 2, act, layout).To(o1)
# Additional data type
quant8 = DataTypeConverter().Identify({
i1: ("TENSOR_QUANT8_ASYMM", 0.5, 0),
w1: ("TENSOR_QUANT8_ASYMM", 0.5, 0),
b1: ("TENSOR_INT32", 0.25, 0),
o1: ("TENSOR_QUANT8_ASYMM", 0.5, 0)
})
quant8_mult_gt_1 = DataTypeConverter().Identify({
i1: ("TENSOR_QUANT8_ASYMM", 0.5, 100),
w1: ("TENSOR_QUANT8_ASYMM", 0.5, 128),
b1: ("TENSOR_INT32", 0.25, 0),
o1: ("TENSOR_QUANT8_ASYMM", 0.1, 80)
})
# Per-channel quantization
channelQuant8 = DataTypeConverter().Identify({
i1: ("TENSOR_QUANT8_ASYMM", 0.25, 100),
w1: ("TENSOR_QUANT8_SYMM_PER_CHANNEL", 0, 0, SymmPerChannelQuantParams(channelDim=0, scales=[0.25, 0.5])),
b1: ("TENSOR_INT32", 0.0, 0, SymmPerChannelQuantParams(channelDim=0, scales=[0.0625, 0.125], hide=True)),
o1: ("TENSOR_QUANT8_ASYMM", 0.5, 80)
})
channelQuant8_mult_gt_1 = DataTypeConverter().Identify({
i1: ("TENSOR_QUANT8_ASYMM", 0.25, 100),
w1: ("TENSOR_QUANT8_SYMM_PER_CHANNEL", 0, 0, SymmPerChannelQuantParams(channelDim=0, scales=[0.25, 0.5])),
b1: ("TENSOR_INT32", 0.0, 0, SymmPerChannelQuantParams(channelDim=0, scales=[0.0625, 0.125], hide=True)),
o1: ("TENSOR_QUANT8_ASYMM", 0.1, 80)
})
Example({
i1: [1, 2, 3, 4],
o1: [-0.5, 0, 1.5, 2, 5.5, 8, 4.5, 6, 8.5, 10,
5.5, 6, 7.5, 8, 23.5, 26, 16.5, 18, 20.5, 22,
14.5, 18, 22.5, 26, 60.5, 70, 40.5, 46, 52.5, 58,
19.5, 22, 25.5, 28, 59.5, 66, 34.5, 38, 42.5, 46,
37.5, 40, 43.5, 46, 101.5, 108, 58.5, 62, 66.5, 70]
}).AddNchw(i1, o1, s1, layout).AddAllActivations(o1, act).AddVariations("relaxed", quant8, quant8_mult_gt_1, channelQuant8, channelQuant8_mult_gt_1, "float16").AddInput(w1, b1)
# TEST 2: TRANSPOSE_CONV2D_LARGE, pad = same, stride = 3, act = relu
i2 = Input("op1", "TENSOR_FLOAT32", "{1, 1, 2, 1}") # input 0
w2 = Parameter("op2", "TENSOR_FLOAT32", "{1, 3, 3, 1}", [9, 5, 6, 9, 8, 5, 3, 1, 4]) # weight
b2 = Parameter("op3", "TENSOR_FLOAT32", "{1}", [-1000]) # bias
s2 = Int32Vector("shape", [1, 3, 4, 1]) # output shape
o2 = Output("op4", "TENSOR_FLOAT32", "{1, 3, 4, 1}") # output
Model().Operation("TRANSPOSE_CONV_2D", i2, w2, b2, s2, 1, 3, 3, 1, layout).To(o2)
# Additional data type
quant8 = DataTypeConverter().Identify({
i2: ("TENSOR_QUANT8_ASYMM", 2.0, 0),
w2: ("TENSOR_QUANT8_ASYMM", 0.25, 128),
b2: ("TENSOR_INT32", 0.5, 0),
o2: ("TENSOR_QUANT8_ASYMM", 20.0, 50)
})
# Per-channel quantization
channelQuant8 = DataTypeConverter().Identify({
i2: ("TENSOR_QUANT8_ASYMM", 2.0, 0),
w2: ("TENSOR_QUANT8_SYMM_PER_CHANNEL", 0, 0, SymmPerChannelQuantParams(channelDim=0, scales=[0.25])),
b2: ("TENSOR_INT32", 0.0, 0, SymmPerChannelQuantParams(channelDim=0, scales=[0.5], hide=True)),
o2: ("TENSOR_QUANT8_ASYMM", 20.0, 50)
})
Example({
i2: [300, 500],
o2: [500., 800., 3500., 1500.,
1400., 500., 3500., 3000.,
0., 200., 500., 0.]
}).AddNchw(i2, o2, s2, layout).AddVariations("relaxed", quant8, channelQuant8, "float16").AddInput(w2, b2)
# TEST 3: TRANSPOSE_CONV2D_SAME, outputShape = [1, 4, 4, 1], pad = same, stride = 1, act = none
i3 = Input("op1", "TENSOR_FLOAT32", "{1, 4, 4, 2}") # input 0
w3 = Parameter("op2", "TENSOR_FLOAT32", "{1, 3, 3, 2}", [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]) # weight
b3 = Parameter("op3", "TENSOR_FLOAT32", "{1}", [0]) # bias
s3 = Int32Vector("shape", [1, 4, 4, 1]) # output shape
o3 = Output("op4", "TENSOR_FLOAT32", "{1, 4, 4, 1}") # output
Model().Operation("TRANSPOSE_CONV_2D", i3, w3, b3, s3, 1, 1, 1, 0, layout).To(o3)
# Additional data type
quant8 = DataTypeConverter().Identify({
i3: ("TENSOR_QUANT8_ASYMM", 0.5, 100),
w3: ("TENSOR_QUANT8_ASYMM", 0.5, 128),
b3: ("TENSOR_INT32", 0.25, 0),
o3: ("TENSOR_QUANT8_ASYMM", 16.0, 0)
})
Example({
i3: [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],
o3: [184, 412, 568, 528,
678, 1347, 1689, 1434,
1494, 2715, 3057, 2442,
1968, 3352, 3652, 2760]
}).AddNchw(i3, o3, s3, layout).AddVariations("relaxed", quant8, "float16").AddInput(w3, b3)
# TEST 4: TRANSPOSE_CONV2D_VALID, outputShape = [1, 6, 6, 1], pad = valid, stride = 1, act = none
i4 = Input("op1", "TENSOR_FLOAT32", "{1, 4, 4, 2}") # input 0
w4 = Parameter("op2", "TENSOR_FLOAT32", "{1, 3, 3, 2}", [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]) # weight
b4 = Parameter("op3", "TENSOR_FLOAT32", "{1}", [0]) # bias
s4 = Int32Vector("shape", [1, 6, 6, 1]) # output shape
o4 = Output("op4", "TENSOR_FLOAT32", "{1, 6, 6, 1}") # output
Model().Operation("TRANSPOSE_CONV_2D", i4, w4, b4, s4, 2, 1, 1, 0, layout).To(o4)
# Additional data type
quant8 = DataTypeConverter().Identify({
i4: ("TENSOR_QUANT8_ASYMM", 0.25, 10),
w4: ("TENSOR_QUANT8_ASYMM", 0.5, 128),
b4: ("TENSOR_INT32", 0.125, 0),
o4: ("TENSOR_QUANT8_ASYMM", 32.0, 80)
})
Example({
i4: [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],
o4: [5, 22, 59, 101, 114, 83,
52, 184, 412, 568, 528, 344,
237, 678, 1347, 1689, 1434, 879,
597, 1494, 2715, 3057, 2442, 1431,
856, 1968, 3352, 3652, 2760, 1548,
689, 1534, 2543, 2729, 2010, 1103]
}).AddNchw(i4, o4, s4, layout).AddVariations("relaxed", quant8, "float16").AddInput(w4, b4)
# TEST 5: TRANSPOSE_CONV2D_EXPLICIT, pad = [1, 2, 2, 1], stride = 1, act = none
i5 = Input("op1", "TENSOR_FLOAT32", "{1, 4, 4, 2}") # input 0
w5 = Parameter("op2", "TENSOR_FLOAT32", "{1, 3, 3, 2}", [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]) # weight
b5 = Parameter("op3", "TENSOR_FLOAT32", "{1}", [0]) # bias
o5 = Output("op4", "TENSOR_FLOAT32", "{1, 3, 3, 1}") # output
Model().Operation("TRANSPOSE_CONV_2D", i5, w5, b5, 1, 2, 2, 1, 1, 1, 0, layout).To(o5)
# Additional data type
quant8 = DataTypeConverter().Identify({
i5: ("TENSOR_QUANT8_ASYMM", 0.5, 100),
w5: ("TENSOR_QUANT8_ASYMM", 0.25, 128),
b5: ("TENSOR_INT32", 0.125, 0),
o5: ("TENSOR_QUANT8_ASYMM", 20.0, 50)
})
Example({
i5: [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],
o5: [678, 1347, 1689,
1494, 2715, 3057,
1968, 3352, 3652]
}).AddNchw(i5, o5, layout).AddVariations("relaxed", quant8, "float16").AddInput(w5, b5)
# TEST 6: zero-sized input, implicit padding
# Use BOX_WITH_NMS_LIMIT op to generate a zero-sized internal tensor for box cooridnates.
p1 = Parameter("scores", "TENSOR_FLOAT32", "{1, 2}", [0.90, 0.10]) # scores
p2 = Parameter("roi", "TENSOR_FLOAT32", "{1, 8}", [1, 1, 10, 10, 0, 0, 10, 10]) # roi
o1 = Output("scoresOut", "TENSOR_FLOAT32", "{0}") # scores out
o2 = Output("classesOut", "TENSOR_INT32", "{0}") # classes out
tmp1 = Internal("roiOut", "TENSOR_FLOAT32", "{0, 4}") # roi out
tmp2 = Internal("batchSplitOut", "TENSOR_INT32", "{0}") # batch split out
model = Model("zero_sized").Operation("BOX_WITH_NMS_LIMIT", p1, p2, [0], 0.3, -1, 0, 0.4, 1.0, 0.3).To(o1, tmp1, o2, tmp2)
# Use ROI_ALIGN op to convert into zero-sized feature map.
i1 = Input("in", "TENSOR_FLOAT32", "{1, 1, 1, 1}")
zero_sized = Internal("featureMap", "TENSOR_FLOAT32", "{0, 2, 2, 1}")
model = model.Operation("ROI_ALIGN", i1, tmp1, tmp2, 2, 2, 2.0, 2.0, 4, 4, layout).To(zero_sized)
# TRANSPOSE_CONV_2D op with numBatches = 0.
w = Parameter("weights", "TENSOR_FLOAT32", "{2, 3, 3, 1}", [1, 3, 5, 7, 9, 11, 9, 7, 5, 2, 4, 6, 8, 10, 12, 10, 8, 6]) # weight
b = Parameter("bias", "TENSOR_FLOAT32", "{2}", [-1.5, -2]) # bias
s = Int32Vector("shape", [0, 5, 5, 2]) # output shape
o3 = Output("out", "TENSOR_FLOAT32", "{0, 5, 5, 2}") # out
model = model.Operation("TRANSPOSE_CONV_2D", zero_sized, w, b, s, 2, 2, 2, 0, layout).To(o3)
quant8 = DataTypeConverter().Identify({
p1: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
p2: ("TENSOR_QUANT16_ASYMM", 0.125, 0),
o1: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
tmp1: ("TENSOR_QUANT16_ASYMM", 0.125, 0),
i1: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
zero_sized: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
w: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
b: ("TENSOR_INT32", 0.01, 0),
o3: ("TENSOR_QUANT8_ASYMM", 0.1, 128)
})
# Create test case with dummy values.
Example({
i1: [1],
o1: [0],
o2: [0],
o3: [0],
}).AddNchw(i1, zero_sized, o3, s, layout).AddVariations("relaxed", quant8, "float16")
# TEST 7: zero-sized input, explicit padding
# Use BOX_WITH_NMS_LIMIT op to generate a zero-sized internal tensor for box cooridnates.
p1 = Parameter("scores", "TENSOR_FLOAT32", "{1, 2}", [0.90, 0.10]) # scores
p2 = Parameter("roi", "TENSOR_FLOAT32", "{1, 8}", [1, 1, 10, 10, 0, 0, 10, 10]) # roi
o1 = Output("scoresOut", "TENSOR_FLOAT32", "{0}") # scores out
o2 = Output("classesOut", "TENSOR_INT32", "{0}") # classes out
tmp1 = Internal("roiOut", "TENSOR_FLOAT32", "{0, 4}") # roi out
tmp2 = Internal("batchSplitOut", "TENSOR_INT32", "{0}") # batch split out
model = Model("zero_sized").Operation("BOX_WITH_NMS_LIMIT", p1, p2, [0], 0.3, -1, 0, 0.4, 1.0, 0.3).To(o1, tmp1, o2, tmp2)
# Use ROI_ALIGN op to convert into zero-sized feature map.
i1 = Input("in", "TENSOR_FLOAT32", "{1, 1, 1, 1}")
zero_sized = Internal("featureMap", "TENSOR_FLOAT32", "{0, 4, 4, 1}")
model = model.Operation("ROI_ALIGN", i1, tmp1, tmp2, 4, 4, 2.0, 2.0, 4, 4, layout).To(zero_sized)
# TRANSPOSE_CONV_2D op with numBatches = 0.
w = Parameter("weights", "TENSOR_FLOAT32", "{1, 3, 3, 1}", [1, 3, 5, 7, 9, 11, 9, 7, 5]) # weight
b = Parameter("bias", "TENSOR_FLOAT32", "{1}", [-1.5]) # bias
o3 = Output("out", "TENSOR_FLOAT32", "{0, 3, 3, 1}") # out
model = model.Operation("TRANSPOSE_CONV_2D", zero_sized, w, b, 1, 2, 2, 1, 1, 1, 0, layout).To(o3)
quant8 = DataTypeConverter().Identify({
p1: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
p2: ("TENSOR_QUANT16_ASYMM", 0.125, 0),
o1: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
tmp1: ("TENSOR_QUANT16_ASYMM", 0.125, 0),
i1: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
zero_sized: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
w: ("TENSOR_QUANT8_ASYMM", 0.1, 128),
b: ("TENSOR_INT32", 0.01, 0),
o3: ("TENSOR_QUANT8_ASYMM", 0.1, 128)
})
# Create test case with dummy values.
Example({
i1: [1],
o1: [0],
o2: [0],
o3: [0],
}).AddNchw(i1, zero_sized, o3, layout).AddVariations("relaxed", quant8, "float16")
# TEST 8: TRANSPOSE_CONV2D_SAME, outputShape = [1, 4, 4, 1], pad = same, stride = 2, act = none
i8 = Input("op1", "TENSOR_FLOAT32", "{1, 2, 2, 1}") # input 0
w8 = Parameter("op2", "TENSOR_FLOAT32", "{1, 1, 1, 1}", [2]) # weight
b8 = Parameter("op3", "TENSOR_FLOAT32", "{1}", [0]) # bias
s8 = Int32Vector("shape", [1, 4, 4, 1]) # output shape
o8 = Output("op4", "TENSOR_FLOAT32", "{1, 4, 4, 1}") # output
Model().Operation("TRANSPOSE_CONV_2D", i8, w8, b8, s8, 1, 2, 2, 0, layout).To(o8)
# Additional data type
quant8 = DataTypeConverter().Identify({
i8: ("TENSOR_QUANT8_ASYMM", 0.5, 100),
w8: ("TENSOR_QUANT8_ASYMM", 0.5, 128),
b8: ("TENSOR_INT32", 0.25, 0),
o8: ("TENSOR_QUANT8_ASYMM", 16.0, 0)
})
Example({
i8: [1, 2, 3, 4],
o8: [2, 0, 4, 0, 0, 0, 0, 0, 6, 0, 8, 0, 0, 0, 0, 0]
}).AddNchw(i8, o8, s8, layout).AddVariations("relaxed", quant8, "float16").AddInput(w8, b8)
|
