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
|
import numpy as np
from numpy.testing import assert_array_equal
from caffe2.python import core, workspace
from caffe2.python.test_util import TestCase
from caffe2.proto import caffe2_pb2
class TestLengthsToShapeOps(TestCase):
def test_lengths_to_shape_ops(self):
workspace.FeedBlob('l', np.array([200, 200, 200], dtype=np.int32))
workspace.RunOperatorOnce(core.CreateOperator(
'LengthsToShape', ['l'], ['s']))
workspace.FeedBlob('res', np.array([3, 200], dtype=np.int32))
assert_array_equal(workspace.FetchBlob('s'), workspace.FetchBlob('res'))
def test_reshape_ops(self):
workspace.FeedBlob('res', np.array([[0, 0, 0, 0]], dtype=np.float32))
workspace.FeedBlob('shape', np.array([1, 4], dtype=np.int32))
workspace.FeedBlob('input', np.zeros((2, 2), dtype=np.float32))
workspace.RunOperatorOnce(core.CreateOperator(
'Reshape', ['input', 'shape'], ['output', 'old_shape']))
assert_array_equal(workspace.FetchBlob('output'),
workspace.FetchBlob('res'))
def test_basic_reshape(self):
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(2, 4))
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(2, 4), arg_shape=False)
def test_missing_dim(self):
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(-1, 8))
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(-1, 8), arg_shape=False)
def test_in_place(self):
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(-1, 8), in_place=True)
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(-1, 8),
in_place=True, arg_shape=False)
def test_zero_dim(self):
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(0, 0, 0),
expected_shape=(4, 2, 1))
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(0, 0, 0),
expected_shape=(4, 2, 1), arg_shape=False)
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(0, 2, 1),
expected_shape=(4, 2, 1))
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(0, 2, 1),
expected_shape=(4, 2, 1), arg_shape=False)
_test_reshape_output_and_gradient(old_shape=(0, 0), new_shape=(0, 0, 0),
expected_shape=(0, 0, 0), arg_shape=False)
def test_zero_dim_and_missing_dim(self):
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(0, -1, 0),
expected_shape=(4, 2, 1))
_test_reshape_output_and_gradient(old_shape=(4, 2, 1), new_shape=(0, -1, 0),
expected_shape=(4, 2, 1), arg_shape=False)
_test_reshape_output_and_gradient(old_shape=(4, 3, 2), new_shape=(-1, 0),
expected_shape=(8, 3))
_test_reshape_output_and_gradient(old_shape=(4, 3, 2), new_shape=(-1, 0),
expected_shape=(8, 3), arg_shape=False)
# empty tensor will just have -1 dim = 0
_test_reshape_output_and_gradient(
old_shape=(2, 0),
new_shape=(-1, 0),
expected_shape=(0, 0),
arg_shape=False
)
def test_backprop(self):
old_shape = (4, 2, 1)
new_shape = (1, 8)
X = np.random.rand(*old_shape).astype(np.float32)
Y = np.random.rand(*new_shape).astype(np.float32)
net = core.Net('net')
net.GivenTensorFill([], 'X', shape=old_shape, values=X.flatten())
net.GivenTensorFill([], 'Y', shape=new_shape, values=Y.flatten())
net.Reshape(['X'], ['X_out', 'old_shape'], shape=new_shape)
net.DotProduct(['X_out', 'Y'], 'Z')
net.AddGradientOperators(['Z'])
workspace.RunNetOnce(net)
Z = workspace.FetchBlob('Z')
X_grad = workspace.FetchBlob('X_grad')
# Check forward computation
np.testing.assert_allclose(
Z.squeeze(), X.reshape(new_shape).dot(Y.T).squeeze(), rtol=1e-5)
# Check the shape of the gradient
np.testing.assert_array_equal(X_grad.shape, X.shape)
# Check the gradient
np.testing.assert_allclose(X_grad, Y.reshape(old_shape), rtol=1e-5)
def test_input_shape_changes(self):
workspace.FeedBlob(
'input_blob',
np.array(np.random.rand(10, 20, 10), dtype=np.float32))
net = core.Net('mynet')
z, _ = net.Reshape('input_blob',
['z_reshape', 'dummy_size'],
shape=(-1, 10))
workspace.CreateNet(net)
workspace.RunNet(net)
workspace.FeedBlob(
'input_blob',
np.array(np.random.rand(10, 40, 10), dtype=np.float32))
workspace.RunNet(net)
def test_nonempty_tensor_gradient(self):
old_shape = [4, 2]
new_shape = [1, 2, -1]
expected_new_shape = [1, 2, 4]
_test_reshape_output_and_gradient(
old_shape=old_shape,
new_shape=new_shape,
expected_shape=expected_new_shape,
expected_gradient=np.ones(shape=old_shape)
)
def test_empty_tensor(self):
old_shape = [4, 0]
new_shape = [1, -1]
expected_new_shape = [1, 0]
_test_reshape_output_and_gradient(
old_shape=old_shape,
new_shape=new_shape,
expected_shape=expected_new_shape,
expected_gradient=np.empty(shape=old_shape)
)
def test_one_dim_empty_tensor_gradient(self):
old_shape = (0,)
new_shape = [1, -1]
expected_new_shape = [1, 0]
_test_reshape_output_and_gradient(
old_shape=old_shape,
new_shape=new_shape,
expected_shape=expected_new_shape,
expected_gradient=np.empty(shape=old_shape)
)
def test_one_dim_and_empty_tensor(self):
old_shape = (0,)
new_shape = [0, -1]
expected_new_shape = [0, 0]
_test_reshape_output_and_gradient(old_shape=old_shape, new_shape=new_shape, expected_shape=expected_new_shape)
def test_scalar_to_tensor(self):
old_shape = ()
new_shape = [1, -1]
expected_new_shape = [1, 1]
_test_reshape_output_and_gradient(old_shape=old_shape, new_shape=new_shape, expected_shape=expected_new_shape)
def _test_reshape_output_and_gradient(
old_shape,
new_shape,
expected_shape=None,
arg_shape=True,
in_place=False,
expected_gradient=None
):
devices = [core.DeviceOption(caffe2_pb2.CPU, 0)]
if workspace.NumGpuDevices() > 0:
devices.append(core.DeviceOption(workspace.GpuDeviceType, 0))
for device_opt in devices:
with core.DeviceScope(device_opt):
if expected_shape is None:
expected_shape = new_shape
net = core.Net('net')
if len(old_shape) == 0:
# scalar, convert to tensor before feeding to blob
X = np.atleast_1d(np.random.rand(*old_shape))
else:
X = np.random.rand(*old_shape).astype(np.float32)
blob_in = 'X'
blob_out = blob_in if in_place else blob_in + '_out'
if arg_shape:
out, _ = net.Reshape([blob_in], [blob_out, 'old_shape'], shape=new_shape)
else:
out, _ = net.Reshape([blob_in, 'new_shape'], [blob_out, 'old_shape'])
workspace.FeedBlob('new_shape', np.asarray(new_shape))
workspace.FeedBlob(blob_in, X)
if expected_gradient is not None:
net.AddGradientOperators([out])
workspace.CreateNet(net)
workspace.RunNetOnce(net)
Y = workspace.FetchBlob(blob_out)
np.testing.assert_allclose(Y, X.reshape(expected_shape))
if expected_gradient is not None:
data_grad = workspace.FetchBlob(blob_in + '_grad')
np.testing.assert_array_equal(data_grad, expected_gradient)
if __name__ == "__main__":
import unittest
unittest.main()
|
