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
|
from caffe2.python import core
import caffe2.python.hypothesis_test_util as hu
import caffe2.python.serialized_test.serialized_test_util as serial
from hypothesis import given, settings
import hypothesis.strategies as st
import numpy as np
import unittest
class TestGroupNormOp(serial.SerializedTestCase):
def group_norm_nchw_ref(self, X, gamma, beta, group, epsilon):
dims = X.shape
N = dims[0]
C = dims[1]
G = group
D = int(C / G)
X = X.reshape(N, G, D, -1)
mu = np.mean(X, axis=(2, 3), keepdims=True)
std = np.sqrt((np.var(X, axis=(2, 3), keepdims=True) + epsilon))
gamma = gamma.reshape(G, D, 1)
beta = beta.reshape(G, D, 1)
Y = gamma * (X - mu) / std + beta
return [Y.reshape(dims), mu.reshape(N, G), (1.0 / std).reshape(N, G)]
def group_norm_nhwc_ref(self, X, gamma, beta, group, epsilon):
dims = X.shape
N = dims[0]
C = dims[-1]
G = group
D = int(C / G)
X = X.reshape(N, -1, G, D)
mu = np.mean(X, axis=(1, 3), keepdims=True)
std = np.sqrt((np.var(X, axis=(1, 3), keepdims=True) + epsilon))
gamma = gamma.reshape(G, D)
beta = beta.reshape(G, D)
Y = gamma * (X - mu) / std + beta
return [Y.reshape(dims), mu.reshape(N, G), (1.0 / std).reshape(N, G)]
@serial.given(
N=st.integers(1, 5), G=st.integers(1, 5), D=st.integers(1, 5),
H=st.integers(2, 5), W=st.integers(2, 5),
epsilon=st.floats(min_value=1e-5, max_value=1e-4),
order=st.sampled_from(["NCHW", "NHWC"]), **hu.gcs)
def test_group_norm_2d(
self, N, G, D, H, W, epsilon, order, gc, dc):
op = core.CreateOperator(
"GroupNorm",
["X", "gamma", "beta"],
["Y", "mean", "inv_std"],
group=G,
epsilon=epsilon,
order=order,
)
C = G * D
if order == "NCHW":
X = np.random.randn(N, C, H, W).astype(np.float32) + 1.0
else:
X = np.random.randn(N, H, W, C).astype(np.float32) + 1.0
gamma = np.random.randn(C).astype(np.float32)
beta = np.random.randn(C).astype(np.float32)
inputs = [X, gamma, beta]
def ref_op(X, gamma, beta):
if order == "NCHW":
return self.group_norm_nchw_ref(X, gamma, beta, G, epsilon)
else:
return self.group_norm_nhwc_ref(X, gamma, beta, G, epsilon)
self.assertReferenceChecks(
device_option=gc,
op=op,
inputs=inputs,
reference=ref_op,
threshold=5e-3,
)
self.assertDeviceChecks(dc, op, inputs, [0, 1, 2])
@given(N=st.integers(1, 5), G=st.integers(1, 3), D=st.integers(2, 3),
T=st.integers(2, 4), H=st.integers(2, 4), W=st.integers(2, 4),
epsilon=st.floats(min_value=1e-5, max_value=1e-4),
order=st.sampled_from(["NCHW", "NHWC"]), **hu.gcs)
def test_group_norm_3d(
self, N, G, D, T, H, W, epsilon, order, gc, dc):
op = core.CreateOperator(
"GroupNorm",
["X", "gamma", "beta"],
["Y", "mean", "inv_std"],
group=G,
epsilon=epsilon,
order=order,
)
C = G * D
if order == "NCHW":
X = np.random.randn(N, C, T, H, W).astype(np.float32) + 1.0
else:
X = np.random.randn(N, T, H, W, C).astype(np.float32) + 1.0
gamma = np.random.randn(C).astype(np.float32)
beta = np.random.randn(C).astype(np.float32)
inputs = [X, gamma, beta]
def ref_op(X, gamma, beta):
if order == "NCHW":
return self.group_norm_nchw_ref(X, gamma, beta, G, epsilon)
else:
return self.group_norm_nhwc_ref(X, gamma, beta, G, epsilon)
self.assertReferenceChecks(
device_option=gc,
op=op,
inputs=inputs,
reference=ref_op,
threshold=5e-3,
)
self.assertDeviceChecks(dc, op, inputs, [0, 1, 2])
@given(N=st.integers(1, 5), G=st.integers(1, 5), D=st.integers(2, 2),
H=st.integers(2, 5), W=st.integers(2, 5),
epsilon=st.floats(min_value=1e-5, max_value=1e-4),
order=st.sampled_from(["NCHW", "NHWC"]), **hu.gcs)
@settings(deadline=10000)
def test_group_norm_grad(
self, N, G, D, H, W, epsilon, order, gc, dc):
op = core.CreateOperator(
"GroupNorm",
["X", "gamma", "beta"],
["Y", "mean", "inv_std"],
group=G,
epsilon=epsilon,
order=order,
)
C = G * D
X = np.arange(N * C * H * W).astype(np.float32)
np.random.shuffle(X)
if order == "NCHW":
X = X.reshape((N, C, H, W))
else:
X = X.reshape((N, H, W, C))
gamma = np.random.randn(C).astype(np.float32)
beta = np.random.randn(C).astype(np.float32)
inputs = [X, gamma, beta]
for i in range(len(inputs)):
self.assertGradientChecks(gc, op, inputs, i, [0])
if __name__ == "__main__":
unittest.main()
|
