pytorch/caffe2/python/operator_test/gather_ops_test.py at master · WeiLinxiao/pytorch · GitHub

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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import numpy as np

from caffe2.python import core, workspace
from hypothesis import given
import caffe2.python.hypothesis_test_util as hu
import caffe2.python.serialized_test.serialized_test_util as serial
import hypothesis.strategies as st
import hypothesis.extra.numpy as hnp

# Basic implementation of gather for axis == 0, shich is lookup of indices
# in the outer dimention. Keeping it for reference here, although is similar
# to more general funciton below.
def ref_gather_axis0():
    def inner(data, ind):
        if ind.size == 0 or data.shape[0] == 0:
            return [np.zeros((0, 10, 20)).astype(np.float32)]
        output = [data[i] for i in ind]
        return [output]
    return inner

# Returns axis-based lookup. We just use numpy take() which handles different
# axis values as we want.
def ref_gather(axis):
    def inner(data, ind):
        if ind.size == 0 or data.shape[axis] == 0:
            shape = list(data.shape)
            shape[0] = 0
            return [np.zeros(tuple(shape)).astype(np.float32)]
        # np.take() does axis lookup same as gather
        output = data.take(ind, axis).astype(np.float32)
        return [output]
    return inner

class TestGatherOps(serial.SerializedTestCase):
    @given(rows_num=st.integers(0, 10000),
           index_num=st.integers(0, 5000),
           **hu.gcs)
    def test_gather_ops(self, rows_num, index_num, gc, dc):
        data = np.random.random((rows_num, 10, 20)).astype(np.float32)

        if rows_num > 0:
            ind = np.random.randint(rows_num, size=(index_num, )).astype('int32')
        else:
            ind = np.random.randint(10, size=(index_num, )).astype('int32')
        op = core.CreateOperator(
            'Gather',
            ['data', 'ind'],
            ['output'])

        self.assertReferenceChecks(gc, op, [data, ind], ref_gather_axis0())
        self.assertDeviceChecks(dc, op, [data, ind], [0])
        return

    # Test axis == 2, this keeps outer dimension but will replace data
    # within axis by lookup of index array (repeated for each outer entry)
    @given(batch_num=st.integers(1, 4000),
        rows_num=st.integers(1, 6),
        index_num=st.integers(1, 20),
        **hu.gcs)
    def test_gather_ops_axis2(self, batch_num, rows_num, index_num, gc, dc):
        data = np.random.random((batch_num, rows_num, 5)).astype(np.float32)
        ind = np.random.randint(5, size=(index_num, )).astype('int32')
        op = core.CreateOperator(
            'Gather',
            ['data', 'ind'],
            ['output'],
            axis=2)

        self.assertReferenceChecks(gc, op, [data, ind], ref_gather(axis=2))
        self.assertDeviceChecks(dc, op, [data, ind], [0])
        return


# Generates data arrays of max dims 10x100x2 and indexing array up to rows_num
@st.composite
def _inputs(draw):
    batch_size = draw(st.integers(2, 10))
    rows_num = draw(st.integers(1, 100))
    block_size = draw(st.integers(1, 2))
    index_num = draw(st.integers(1, 10))
    return (
        draw(hnp.arrays(
            np.float32,
            (batch_size, rows_num, block_size),
            elements=st.floats(-10.0, 10.0),
        )),
        draw(hnp.arrays(
            np.int32,
            (index_num, 1),
            elements=st.integers(0, rows_num - 1),
        )),
    )

class TestBatchGatherOps(hu.HypothesisTestCase):
    @given(inputs=_inputs(),
           **hu.gcs)
    def test_batch_gather_ops(self, inputs, gc, dc):
        data, ind = inputs
        op = core.CreateOperator(
            'BatchGather',
            ['data', 'ind'],
            ['output'])
        self.assertReferenceChecks(gc, op, [data, ind], ref_gather(axis=1))
        self.assertGradientChecks(gc, op, [data, ind], 0, [0])


class TestGatherFused8BitRowwise(hu.HypothesisTestCase):
    @given(rows_num=st.integers(1, 10000),
           cols_num=st.integers(1, 128),
           index_num=st.integers(0, 5000),
           **hu.gcs)
    def test_batch_gather_ops(self, rows_num, cols_num, index_num, gc, dc):
        data = np.random.random((rows_num, cols_num)).astype(np.float32)
        ind = np.random.randint(rows_num, size=(index_num, )).astype('int32')

        net = core.Net("bench")

        quantized_data = net.FloatToFused8BitRowwiseQuantized(
            'data', 'quantized_data')
        dequantized_data = net.Fused8BitRowwiseQuantizedToFloat(
            quantized_data, 'dequantized_data')

        net.Gather(
            [dequantized_data, 'ind'], 'gather_reference')
        net.GatherFused8BitRowwise(
            [quantized_data, 'ind'], 'gather_quantized')

        workspace.FeedBlob('data', data)
        workspace.FeedBlob('ind', ind)
        workspace.CreateNet(net)
        workspace.RunNetOnce(net)

        gather_reference = workspace.FetchBlob('gather_reference')
        gather_quantized = workspace.FetchBlob('gather_quantized')
        np.testing.assert_array_almost_equal(gather_reference, gather_quantized)


if __name__ == "__main__":
    import unittest
    unittest.main()