pytorch

1
# Owner(s): ["module: fx"]
2

3
import copy
4
import unittest
5
from collections import defaultdict
6

7
import torch
8
import torch.fx as fx
9
from torch._dynamo.source import LocalSource
10
from torch.fx.experimental.shape_inference.infer_shape import infer_shape
11
from torch.fx.experimental.shape_inference.infer_symbol_values import (
12
    infer_symbol_values,
13
)
14
from torch.fx.experimental.symbolic_shapes import DimDynamic, ShapeEnv
15

16

17
class TestShapeInference(unittest.TestCase):
18
    def test_infer_symbol_values(self):
19
        def mksym(shape_env, value, source, dynamic_dim) -> None:
20
            return shape_env.create_symintnode(
21
                shape_env.create_symbol(
22
                    value,
23
                    source=source,
24
                    dynamic_dim=dynamic_dim,
25
                ),
26
                hint=value,
27
                source=source,
28
            )
29

30
        shape_env = ShapeEnv()
31
        N = 8
32
        sample = {f"s{i}": 2 for i in range(N)}
33
        init_symints = [
34
            mksym(shape_env, v, LocalSource(k), DimDynamic.DYNAMIC)
35
            for k, v in sample.items()
36
        ]
37
        symints = copy.deepcopy(init_symints)
38
        symbol_to_idx_dict = {f"s{i}": i for i in range(N)}
39
        padding_constraints = defaultdict(list)
40

41
        # prepare constraints strings
42
        constraints = []
43
        constraints.append(
44
            "The size of tensor a (s1) must match the size of tensor b (1773) at non-singleton dimension 1)"
45
        )
46
        constraints.append(
47
            "Expected size for first two dimensions of batch2 tensor to be: [s0, (s2//2) + 12] but got: [s0, 120]."
48
        )
49
        constraints.append("shape '[s0, -1, 32]' is invalid for input of size s0*s3")
50
        constraints.append(
51
            "a and b must have same reduction dim, but got [32*s0, s3] X [20, 15]."
52
        )
53
        constraints.append(
54
            "a and b must have same reduction dim, but got [s0, s4 + 1568] X [5728, 1024]."
55
        )
56
        constraints.append(
57
            "Expected size for first two dimensions of batch2 tensor to be: [s0, 40] but got: [s0, s5]."
58
        )
59
        constraints.append(
60
            "shape '[s0, -1, 32]' is invalid for input of size s0*s6 + 1344*s0"
61
        )
62
        constraints.append(
63
            "shape '[-1, 47]' is invalid for input of size 32*s0*s6 + 1344*s0"
64
        )
65
        constraints.append(
66
            "Expected size for first two dimensions of batch2 tensor to be: [s0, 47*s6] but got: [s0*s6, 47]."
67
        )
68
        constraints.append("Split sizes add up to 4258 but got the tensor's size of s7")
69

70
        for constraint in constraints:
71
            infer_symbol_values(
72
                symints,
73
                init_symints,
74
                symbol_to_idx_dict,
75
                padding_constraints,
76
                constraint,
77
            )
78

79
        self.assertEqual(symints[1], 1773)
80
        self.assertEqual(symints[2], 216)
81
        self.assertEqual(symints[3], 640)
82
        self.assertEqual(symints[4], 4160)
83
        self.assertEqual(symints[5], 40)
84
        self.assertEqual(symints[6], 160)
85
        self.assertEqual(symints[7], 4258)
86

87
    def test_infer_shape(self):
88
        class TestModule(torch.nn.Module):
89
            def __init__(self) -> None:
90
                super().__init__()
91
                self.w_1 = torch.empty([256, 328])
92
                self.b_1 = torch.empty([256])
93
                self.w_2 = torch.empty([328, 256])
94
                self.b_2 = torch.empty([328])
95

96
            def forward(self, x):
97
                l_1 = torch.nn.functional.linear(x, self.w_1, bias=self.b_1)
98
                s_1 = torch.sigmoid(l_1)
99
                l_2 = torch.nn.functional.linear(s_1, self.w_2, bias=self.b_2)
100
                t_1 = torch.tanh(l_2)
101
                return t_1
102

103
        def generate_graph_module(model):
104
            gm = fx.symbolic_trace(model)
105
            return gm
106

107
        m = TestModule()
108
        gm = generate_graph_module(m)
109
        input_tensors = [torch.randn(1, 1)]
110
        infer_shape(gm, input_tensors)
111