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
|
import collections
import caffe2.python.hypothesis_test_util as hu
import hypothesis.strategies as st
import numpy as np
from caffe2.python import core, dyndep, workspace
from caffe2.quantization.server.dnnlowp_test_utils import check_quantized_results_close
from hypothesis import assume, given
dyndep.InitOpsLibrary("//caffe2/caffe2/quantization/server:dnnlowp_ops")
workspace.GlobalInit(["caffe2", "--caffe2_omp_num_threads=11"])
class DNNLowPOpPoolTest(hu.HypothesisTestCase):
@given(
stride=st.integers(1, 3),
pad=st.integers(0, 3),
kernel=st.integers(1, 5),
size=st.integers(1, 20),
input_channels=st.integers(1, 3),
batch_size=st.integers(1, 3),
order=st.sampled_from(["NCHW", "NHWC"]),
in_quantized=st.booleans(),
**hu.gcs_cpu_only
)
def test_dnnlowp_max_pool(
self,
stride,
pad,
kernel,
size,
input_channels,
batch_size,
order,
in_quantized,
gc,
dc,
):
assume(kernel <= size)
assume(pad < kernel)
C = input_channels
N = batch_size
H = W = size
min_ = -10
max_ = 20
if order == "NCHW":
X = np.round(np.random.rand(N, C, H, W) * (max_ - min_) + min_)
elif order == "NHWC":
X = np.round(np.random.rand(N, H, W, C) * (max_ - min_) + min_)
X = X.astype(np.float32)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("MaxPool", ""),
("MaxPool", "DNNLOWP"),
("Int8MaxPool", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
if do_quantize:
quantize = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize])
max_pool = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X"],
["Y_q" if engine == "DNNLOWP" else "Y"],
stride=stride,
kernel=kernel,
pad=pad,
order=order,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([max_pool])
if engine == "DNNLOWP":
dequantize = core.CreateOperator(
"Dequantize", ["Y_q"], ["Y"], engine=engine, device_option=gc
)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(), op_type=op_type, engine=engine)
)
# Y_i = max(X_j) so the only error is in quantization of inputs
check_quantized_results_close(outputs, ref=X)
@given(
ndim=st.integers(2, 3),
stride=st.integers(1, 1),
pad=st.integers(0, 0),
kernel=st.integers(1, 5),
size=st.integers(2, 2),
input_channels=st.integers(1, 1),
batch_size=st.integers(2, 2),
order=st.sampled_from(["NCHW", "NHWC"]),
in_quantized=st.booleans(),
**hu.gcs_cpu_only
)
def test_dnnlowp_average_pool(
self,
ndim,
stride,
pad,
kernel,
size,
input_channels,
batch_size,
order,
in_quantized,
gc,
dc,
):
kernel = 2 # Only kernel size 2 is supported
assume(kernel <= size)
assume(pad < kernel)
C = input_channels
N = batch_size
strides = (stride,) * ndim
pads = (pad,) * (ndim * 2)
kernels = (kernel,) * ndim
sizes = (size,) * ndim
# X has scale 1, so no input quantization error
min_ = -100
max_ = min_ + 255
if order == "NCHW":
X = np.round(np.random.rand(*((N, C) + sizes)) * (max_ - min_) + min_)
X = X.astype(np.float32)
X[(0,) * (ndim + 2)] = min_
X[(0,) * (ndim + 1) + (1,)] = max_
elif order == "NHWC":
X = np.round(np.random.rand(*((N,) + sizes + (C,))) * (max_ - min_) + min_)
X = X.astype(np.float32)
X[(0,) * (ndim + 2)] = min_
X[(0, 1) + (0,) * ndim] = max_
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("AveragePool", ""),
("AveragePool", "DNNLOWP"),
("Int8AveragePool", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
if do_quantize:
quantize = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize])
max_pool = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X"],
["Y_q" if engine == "DNNLOWP" else "Y"],
strides=strides,
kernels=kernels,
pads=pads,
order=order,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([max_pool])
if engine == "DNNLOWP":
dequantize = core.CreateOperator(
"Dequantize", ["Y_q"], ["Y"], engine=engine, device_option=gc
)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(), op_type=op_type, engine=engine)
)
check_quantized_results_close(outputs)
|
