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
|
import numpy as np
from hypothesis import given, assume, settings
import hypothesis.strategies as st
from caffe2.python import core, workspace
import caffe2.python.hypothesis_test_util as hu
import caffe2.python.mkl_test_util as mu
import caffe2.python.serialized_test.serialized_test_util as serial
from scipy.stats import norm
import unittest
class TestActivations(serial.SerializedTestCase):
@given(X=hu.tensor(), in_place=st.booleans(),
engine=st.sampled_from(["", "CUDNN"]), **mu.gcs)
@settings(deadline=10000)
def test_relu(self, X, in_place, engine, gc, dc):
if gc == mu.mkl_do:
in_place = False
op = core.CreateOperator(
"Relu",
["X"],
["X"] if in_place else ["Y"],
engine=engine,
)
def relu_ref(X):
return [np.maximum(X, 0.0)]
# go away from the origin point to avoid kink problems
X += 0.02 * np.sign(X)
X[X == 0.0] += 0.02
self.assertReferenceChecks(gc, op, [X], relu_ref, ensure_outputs_are_inferred=True)
self.assertDeviceChecks(dc, op, [X], [0])
self.assertGradientChecks(gc, op, [X], 0, [0], ensure_outputs_are_inferred=True)
@given(N=st.integers(1, 10), M=st.integers(1, 10), in_place=st.booleans(),
**hu.gcs)
def test_relu_empty_input(self, N, M, in_place, gc, dc):
op = core.CreateOperator(
"Relu",
["X"],
["X"] if in_place else ["Y"],
)
def relu_ref(X):
return [np.maximum(X, 0.0)]
X = np.random.randn(0, N, M).astype(np.float32)
self.assertReferenceChecks(gc, op, [X], relu_ref, ensure_outputs_are_inferred=True)
self.assertDeviceChecks(dc, op, [X], [0])
self.assertGradientChecks(gc, op, [X], 0, [0], ensure_outputs_are_inferred=True)
@unittest.skipIf(not workspace.has_gpu_support,
"Relu for float16 can only run on GPU now.")
@given(X=hu.tensor(dtype=np.float16), in_place=st.booleans(),
engine=st.sampled_from(["", "CUDNN"]), **hu.gcs)
def test_relu_fp16(self, X, in_place, engine, gc, dc):
# fp16 is only supported on CUDA/HIP
assume(core.IsGPUDeviceType(gc.device_type))
op = core.CreateOperator(
"Relu",
["X"],
["X"] if in_place else ["Y"],
engine=engine,
)
def relu_ref(X):
return [np.maximum(X, 0.0)]
def relu_grad_ref(g_out, outputs, fwd_inputs):
dY = g_out
[Y] = outputs
dX = dY
dX[Y == 0] = 0
return [dX]
# go away from the origin point to avoid kink problems
X += 0.02 * np.sign(X)
X[X == 0.0] += 0.02
self.assertReferenceChecks(
gc,
op,
[X],
relu_ref,
output_to_grad="X" if in_place else "Y",
grad_reference=relu_grad_ref)
@serial.given(X=hu.tensor(elements=hu.floats(-3.0, 3.0)),
n=hu.floats(min_value=0.5, max_value=2.0),
in_place=st.booleans(), **hu.gcs)
def test_relu_n(self, X, n, in_place, gc, dc):
op = core.CreateOperator(
"ReluN",
["X"],
["X"] if in_place else ["Y"],
n=n,
)
def relu_n_ref(X):
return [np.minimum(np.maximum(X, 0.0), n)]
# go away from 0 and n to avoid kink problems
X += 0.04 * np.sign(X)
X[X == 0.0] += 0.04
X -= n
X += 0.02 * np.sign(X)
X[X == 0.0] -= 0.02
X += n
self.assertReferenceChecks(gc, op, [X], relu_n_ref, ensure_outputs_are_inferred=True)
self.assertDeviceChecks(dc, op, [X], [0])
self.assertGradientChecks(gc, op, [X], 0, [0], stepsize=0.005,
ensure_outputs_are_inferred=True)
@serial.given(X=hu.tensor(),
alpha=hu.floats(min_value=0.1, max_value=2.0),
in_place=st.booleans(), engine=st.sampled_from(["", "CUDNN"]),
**hu.gcs)
def test_elu(self, X, alpha, in_place, engine, gc, dc):
op = core.CreateOperator(
"Elu",
["X"],
["X"] if in_place else ["Y"],
alpha=alpha,
engine=engine,
)
def elu_ref(X):
Y = X
Y[X < 0] = alpha * (np.exp(X[X < 0]) - 1.0)
return [Y]
# go away from the origin point to avoid kink problems
X += 0.04 * np.sign(X)
X[X == 0.0] += 0.04
self.assertReferenceChecks(gc, op, [X], elu_ref, ensure_outputs_are_inferred=True)
self.assertDeviceChecks(dc, op, [X], [0])
self.assertGradientChecks(gc, op, [X], 0, [0], stepsize=1e-2, ensure_outputs_are_inferred=True)
@given(X=hu.tensor(min_dim=4, max_dim=4),
alpha=hu.floats(min_value=0.1, max_value=2.0),
inplace=st.booleans(),
shared=st.booleans(),
order=st.sampled_from(["NCHW", "NHWC"]),
seed=st.sampled_from([20, 100]),
**hu.gcs)
@settings(deadline=10000)
def test_prelu(self, X, alpha, inplace, shared, order, seed, gc, dc):
np.random.seed(seed)
W = np.random.randn(
X.shape[1] if order == "NCHW" else X.shape[3]).astype(np.float32)
if shared:
W = np.random.randn(1).astype(np.float32)
# go away from the origin point to avoid kink problems
X += 0.04 * np.sign(X)
X[X == 0.0] += 0.04
def prelu_ref(X, W):
Y = X.copy()
W = W.reshape(1, -1, 1, 1) if order == "NCHW" \
else W.reshape(1, 1, 1, -1)
assert len(X.shape) == 4
neg_indices = X <= 0
assert len(neg_indices.shape) == 4
assert X.shape == neg_indices.shape
Y[neg_indices] = (Y * W)[neg_indices]
return (Y,)
op = core.CreateOperator(
"PRelu", ["X", "W"], ["Y" if not inplace else "X"],
alpha=alpha, order=order)
self.assertReferenceChecks(gc, op, [X, W], prelu_ref, ensure_outputs_are_inferred=True)
# Check over multiple devices
self.assertDeviceChecks(dc, op, [X, W], [0])
if not inplace:
# Gradient check wrt X
self.assertGradientChecks(gc, op, [X, W], 0, [0], stepsize=1e-2, ensure_outputs_are_inferred=True)
# Gradient check wrt W
self.assertGradientChecks(gc, op, [X, W], 1, [0], stepsize=1e-2, ensure_outputs_are_inferred=True)
@serial.given(X=hu.tensor(),
alpha=hu.floats(min_value=0.1, max_value=2.0),
inplace=st.booleans(),
**hu.gcs)
def test_leaky_relu(self, X, alpha, inplace, gc, dc):
# go away from the origin point to avoid kink problems
X += 0.04 * np.sign(X)
X[X == 0.0] += 0.04
def leaky_relu_ref(X):
Y = X.copy()
neg_indices = X <= 0
Y[neg_indices] = Y[neg_indices] * alpha
return (Y,)
op = core.CreateOperator(
"LeakyRelu",
["X"], ["Y" if not inplace else "X"],
alpha=alpha)
self.assertReferenceChecks(gc, op, [X], leaky_relu_ref,
ensure_outputs_are_inferred=True)
# Check over multiple devices
self.assertDeviceChecks(dc, op, [X], [0])
@given(X=hu.tensor(),
inplace=st.booleans(),
**hu.gcs)
def test_leaky_relu_default(self, X, inplace, gc, dc):
# go away from the origin point to avoid kink problems
X += 0.04 * np.sign(X)
X[X == 0.0] += 0.04
def leaky_relu_ref(X):
Y = X.copy()
neg_indices = X <= 0
Y[neg_indices] = Y[neg_indices] * 0.01
return (Y,)
op = core.CreateOperator(
"LeakyRelu",
["X"], ["Y" if not inplace else "X"])
self.assertReferenceChecks(gc, op, [X], leaky_relu_ref)
# Check over multiple devices
self.assertDeviceChecks(dc, op, [X], [0])
@given(X=hu.tensor(),
fast_gelu=st.booleans(),
**hu.gcs)
@settings(deadline=10000)
def test_gelu(self, X, fast_gelu, gc, dc):
op = core.CreateOperator(
"Gelu",
["X"],
["Y"],
fast_gelu=fast_gelu,
)
def gelu_ref(X):
return (X * norm.cdf(X),)
tol = 1e-3 if fast_gelu else 1e-4
self.assertReferenceChecks(gc, op, [X], gelu_ref, threshold=tol,
ensure_outputs_are_inferred=True)
self.assertDeviceChecks(dc, op, [X], [0])
self.assertGradientChecks(gc, op, [X], 0, [0],
ensure_outputs_are_inferred=True)
@given(n=st.integers(0, 6), m=st.integers(4, 6),
seed=st.integers(0, 1000), **hu.gcs_cpu_only)
def test_mish(self, n, m, gc, dc, seed):
np.random.seed(seed)
X = np.random.rand(n, m).astype(np.float32)
def mish_ref(X):
return (X * np.tanh(np.log1p(np.exp(X))),)
op = core.CreateOperator(
"Mish",
["X"],
["Y"]
)
self.assertReferenceChecks(
device_option=gc,
op=op,
inputs=[X],
reference=mish_ref,
ensure_outputs_are_inferred=True,
)
self.assertGradientChecks(
gc, op, [X], 0, [0], ensure_outputs_are_inferred=True)
if __name__ == "__main__":
unittest.main()
|
