pytorch

1

2

3
import caffe2.python.hypothesis_test_util as hu
4
import caffe2.python.serialized_test.serialized_test_util as serial
5
import hypothesis.strategies as st
6
import numpy as np
7
from caffe2.python import core
8

9

10
class ChannelShuffleOpsTest(serial.SerializedTestCase):
11
    def _channel_shuffle_nchw_ref(self, X, group):
12
        dims = X.shape
13
        N = dims[0]
14
        C = dims[1]
15
        G = group
16
        K = int(C / G)
17
        X = X.reshape(N, G, K, np.prod(dims[2:]))
18
        Y = np.transpose(X, axes=(0, 2, 1, 3))
19
        return [Y.reshape(dims)]
20

21
    def _channel_shuffle_nhwc_ref(self, X, group):
22
        dims = X.shape
23
        N = dims[0]
24
        C = dims[-1]
25
        G = group
26
        K = int(C / G)
27
        X = X.reshape(N, np.prod(dims[1:-1]), G, K)
28
        Y = np.transpose(X, axes=(0, 1, 3, 2))
29
        return [Y.reshape(dims)]
30

31
    @serial.given(
32
        N=st.integers(0, 5),
33
        G=st.integers(1, 5),
34
        K=st.integers(1, 5),
35
        H=st.integers(1, 5),
36
        W=st.integers(1, 5),
37
        order=st.sampled_from(["NCHW", "NHWC"]),
38
        **hu.gcs
39
    )
40
    def test_channel_shuffle(self, N, G, K, H, W, order, gc, dc):
41
        C = G * K
42
        if order == "NCHW":
43
            X = np.random.randn(N, C, H, W).astype(np.float32)
44
        else:
45
            X = np.random.randn(N, H, W, C).astype(np.float32)
46

47
        op = core.CreateOperator("ChannelShuffle", ["X"], ["Y"], group=G, order=order)
48

49
        def channel_shuffle_ref(X):
50
            if order == "NCHW":
51
                return self._channel_shuffle_nchw_ref(X, G)
52
            else:
53
                return self._channel_shuffle_nhwc_ref(X, G)
54

55
        self.assertReferenceChecks(gc, op, [X], channel_shuffle_ref)
56
        self.assertGradientChecks(gc, op, [X], 0, [0])
57
        self.assertDeviceChecks(dc, op, [X], [0])
58