test_pytorch_onnx_shape_inference.py

# Owner(s): ["module: onnx"]

import io

import numpy as np

import onnx
import pytorch_test_common
from pytorch_test_common import skipIfUnsupportedMinOpsetVersion

import torch
from torch.onnx import _constants, utils
from torch.onnx._globals import GLOBALS
from torch.onnx._internal import jit_utils
from torch.testing._internal import common_utils


def expect_tensor(scalar_type, shape=None):
    def verify(actual_type):
        np.testing.assert_equal(actual_type.scalarType(), scalar_type)
        # if shape is not None:
        #     np.testing.assert_equal(actual_type.sizes(), shape)
        if shape is not None:
            np.testing.assert_equal(actual_type.varyingSizes(), shape)

    return verify


def as_graphcontext(graph: torch.Graph) -> jit_utils.GraphContext:
    return jit_utils.GraphContext(
        graph=graph,
        block=graph.block(),
        opset=_constants.ONNX_TORCHSCRIPT_EXPORTER_MAX_OPSET,
        original_node=None,  # type: ignore[arg-type]
        params_dict={},
        env={},
        values_in_env=set(),
    )


def g_op(graph: torch.Graph, op_name: str, *args, **kwargs):
    return as_graphcontext(graph).op(op_name, *args, **kwargs)


class TestONNXShapeInference(pytorch_test_common.ExportTestCase):
    def setUp(self):
        self.opset_version = _constants.ONNX_TORCHSCRIPT_EXPORTER_MAX_OPSET
        GLOBALS.export_onnx_opset_version = self.opset_version

    def run_test(self, g, n, type_assertion_funcs):
        if not isinstance(type_assertion_funcs, list):
            type_assertion_funcs = [type_assertion_funcs]

        torch._C._jit_pass_onnx_graph_shape_type_inference(g, {}, self.opset_version)
        for out, type_assertion_func in zip(n.outputs(), type_assertion_funcs):
            type_assertion_func(out.type())

    def create_empty_graph(self):
        g = torch._C.Graph()
        # kick off initialization for ConstantMap.
        torch._C._jit_pass_onnx_graph_shape_type_inference(g, {}, self.opset_version)
        return g

    def insert_tensor_constant(self, g, tensor):
        return g_op(g, "Constant", value_t=tensor)

    def test_cast(self):
        # Test cast with input of unknown scalar type.
        g = self.create_empty_graph()
        input = g.addInput()
        cast_out = g_op(g, "Cast", input, to_i=1)
        self.run_test(g, cast_out.node(), expect_tensor("Float"))

    def test_constant_of_shape(self):
        # Test ConstantOfShape with input of onnx::Shape node.
        g = self.create_empty_graph()
        constant = self.insert_tensor_constant(g, torch.ones(1, 2, 3, 4))
        shape = g_op(g, "Shape", constant)
        constant_of_shape = g_op(
            g, "ConstantOfShape", shape, value_t=torch.tensor([2.0])
        )
        self.run_test(
            g, constant_of_shape.node(), expect_tensor("Float", shape=(1, 2, 3, 4))
        )

    def test_constant_of_shape_static(self):
        # Test ConstantOfShape with input of prim::ListConstruct of static tensor
        rank = 4
        g = self.create_empty_graph()
        constants = [
            self.insert_tensor_constant(g, torch.tensor(i + 1)) for i in range(rank)
        ]
        shape = g_op(g, "prim::ListConstruct", *constants)
        shape.setType(torch._C.ListType.ofInts())
        constant_of_shape = g_op(
            g, "ConstantOfShape", shape, value_t=torch.tensor([2.0])
        )
        self.run_test(
            g, constant_of_shape.node(), expect_tensor("Float", shape=(1, 2, 3, 4))
        )

    def test_constant_of_shape_dynamic(self):
        # Test ConstantOfShape with input of prim::ListConstruct of dynamic tensor
        rank = 4
        g = self.create_empty_graph()
        inputs = [g.addInput() for i in range(rank)]
        shape = g_op(g, "prim::ListConstruct", *inputs)
        shape.setType(torch._C.ListType.ofInts())
        constant_of_shape = g_op(
            g, "ConstantOfShape", shape, value_t=torch.tensor([2.0])
        )
        self.run_test(
            g,
            constant_of_shape.node(),
            expect_tensor("Float", shape=(None, None, None, None)),
        )

    def test_gather_dynamic_index(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(
            input.type().with_dtype(torch.float).with_sizes([None, 3, 16, 16])
        )
        indices = g.addInput()
        indices.setType(indices.type().with_dtype(torch.int64).with_sizes([None]))
        output = g_op(g, "Gather", input, indices, axis_i=1)
        self.run_test(
            g, output.node(), expect_tensor("Float", shape=([None, None, 16, 16]))
        )

    def test_gather_scalar_index(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(
            input.type().with_dtype(torch.float).with_sizes([None, 3, 16, 16])
        )
        indices = self.insert_tensor_constant(g, torch.tensor(1))
        output = g_op(g, "Gather", input, indices, axis_i=1)
        self.run_test(g, output.node(), expect_tensor("Float", shape=([None, 16, 16])))

    def test_reshape(self):
        g = self.create_empty_graph()
        constant = self.insert_tensor_constant(g, torch.ones(2, 16, 5, 5))
        constant_2 = self.insert_tensor_constant(g, torch.tensor([2, 0, -1]))
        shape = g_op(g, "Reshape", constant, constant_2)
        self.run_test(g, shape.node(), expect_tensor("Float", shape=(2, 16, 25)))

        g = self.create_empty_graph()
        constant = self.insert_tensor_constant(g, torch.ones(2, 16, 5, 4))
        constant_2 = self.insert_tensor_constant(g, torch.tensor([-1, 0, 4]))
        shape = g_op(g, "Reshape", constant, constant_2)
        self.run_test(g, shape.node(), expect_tensor("Float", shape=(10, 16, 4)))

        g = self.create_empty_graph()
        constant = self.insert_tensor_constant(g, torch.ones(2, 16, 5, 4))
        constant_2 = self.insert_tensor_constant(g, torch.tensor([-1, 0, 0]))
        shape = g_op(g, "Reshape", constant, constant_2)
        self.run_test(g, shape.node(), expect_tensor("Float", shape=(8, 16, 5)))

    def test_reshape_symbolic(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_sizes([None, None, 2, 8]))
        constant = self.insert_tensor_constant(g, torch.tensor([0, 0, -1]))
        output = g_op(g, "Reshape", input, constant)
        self.run_test(g, output.node(), expect_tensor(None, shape=(None, None, 16)))

    @skipIfUnsupportedMinOpsetVersion(14)
    def test_reshape_allowzero(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_sizes([3, 4, 0]))
        constant = self.insert_tensor_constant(g, torch.tensor([0, 4, 3]))
        output = g_op(g, "Reshape", input, constant, allowzero_i=1)
        self.run_test(g, output.node(), expect_tensor(None, shape=(0, 4, 3)))

    def test_slice(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_sizes([None, None]))
        start_input = g.addInput()
        start_input.setType(start_input.type().with_sizes([None]))
        end = self.insert_tensor_constant(g, torch.tensor([3]))
        axis = self.insert_tensor_constant(g, torch.tensor([0]))
        step = self.insert_tensor_constant(g, torch.tensor([1]))
        slice = g_op(g, "Slice", input, start_input, end, axis, step)
        self.run_test(g, slice.node(), expect_tensor(None, shape=(None, None)))

    def test_slice_with_dynamic_start_index(self):
        g = self.create_empty_graph()
        input = self.insert_tensor_constant(g, torch.ones(2, 3, 4, 5))
        start_input = g.addInput()
        start_input.setType(start_input.type().with_sizes([2]))
        end = self.insert_tensor_constant(g, torch.tensor([3, 4]))
        axis = self.insert_tensor_constant(g, torch.tensor([1, -1]))
        slice = g_op(g, "Slice", input, start_input, end, axis)
        self.run_test(g, slice.node(), expect_tensor(None, shape=(2, None, 4, None)))

    def test_broadcast_matmul(self):
        g = self.create_empty_graph()
        constant = self.insert_tensor_constant(g, torch.ones(5, 1, 2))
        constant_2 = self.insert_tensor_constant(g, torch.ones(3, 1, 2, 1))
        shape = g_op(g, "MatMul", constant, constant_2)
        self.run_test(g, shape.node(), expect_tensor("Float", shape=(3, 5, 1, 1)))

        # test when first input is of rank 1
        g = self.create_empty_graph()
        constant = self.insert_tensor_constant(g, torch.ones(2))
        constant_2 = self.insert_tensor_constant(g, torch.ones(3, 1, 2, 1))
        shape = g_op(g, "MatMul", constant, constant_2)
        self.run_test(g, shape.node(), expect_tensor("Float", shape=(3, 1, 1)))

        # test when second input is of rank 1
        g = self.create_empty_graph()
        constant = self.insert_tensor_constant(g, torch.ones(5, 1, 2))
        constant_2 = self.insert_tensor_constant(g, torch.ones(2))
        shape = g_op(g, "MatMul", constant, constant_2)
        self.run_test(g, shape.node(), expect_tensor("Float", shape=(5, 1)))

        # test when both inputs are of rank 1
        g = self.create_empty_graph()
        constant = self.insert_tensor_constant(g, torch.ones(2))
        constant_2 = self.insert_tensor_constant(g, torch.ones(2))
        shape = g_op(g, "MatMul", constant, constant_2)
        self.run_test(g, shape.node(), expect_tensor("Float", shape=()))

    def test_expand(self):
        g = self.create_empty_graph()
        input = g.addInput()
        constant = self.insert_tensor_constant(g, torch.ones(2, 4))
        input.setType(constant.type().with_sizes([None, None]))
        shape = g_op(g, "Shape", input)
        expand = g_op(g, "Expand", constant, shape)
        self.run_test(g, expand.node(), expect_tensor("Float", shape=(None, None)))

    def test_pad(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_dtype(torch.float).with_sizes([3, 320, 100]))
        constant = self.insert_tensor_constant(g, torch.ones(6, dtype=torch.long))
        none = g_op(g, "prim::Constant").setType(torch.NoneType.get())
        pad = g_op(g, "Pad", input, constant, none, mode_s="constant")
        self.run_test(g, pad.node(), expect_tensor("Float", shape=(5, 322, 102)))

    def test_pad_with_dynamic_input_shape(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_dtype(torch.float).with_sizes([3, None, None]))
        constant = self.insert_tensor_constant(g, torch.ones(6, dtype=torch.long))
        none = g_op(g, "prim::Constant").setType(torch.NoneType.get())
        pad = g_op(g, "Pad", input, constant, none, mode_s="constant")
        self.run_test(g, pad.node(), expect_tensor("Float", shape=(5, None, None)))

    def test_pad_with_dynamic_pad_size(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_dtype(torch.float).with_sizes([3, 320, 100]))
        pad_size = g.addInput()
        pad_size.setType(pad_size.type().with_dtype(torch.long).with_sizes([6]))
        none = g_op(g, "prim::Constant").setType(torch.NoneType.get())
        pad = g_op(g, "Pad", input, pad_size, none, mode_s="constant")
        self.run_test(g, pad.node(), expect_tensor("Float", shape=(None, None, None)))

    def test_resize(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_dtype(torch.float).with_sizes([4, 32, 64, 64]))
        none = g_op(g, "prim::Constant").setType(torch.NoneType.get())
        scales = self.insert_tensor_constant(
            g, torch.tensor([1, 1, 2, 2], dtype=torch.float)
        )
        resize = g_op(
            g,
            "Resize",
            input,
            none,
            scales,
            coordinate_transformation_mode_s="align_corners",
            cubic_coeff_a_f=-0.75,
            mode_s="linear",
            nearest_mode_s="floor",
        )
        self.run_test(g, resize.node(), expect_tensor("Float", shape=(4, 32, 128, 128)))

    def test_resize_after_concat(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_dtype(torch.float).with_sizes([4, 32, 64, 64]))
        none = g_op(g, "prim::Constant").setType(torch.NoneType.get())
        scale_1 = self.insert_tensor_constant(
            g, torch.tensor([1, 1], dtype=torch.float)
        )
        scale_2 = self.insert_tensor_constant(
            g, torch.tensor([2, 2], dtype=torch.float)
        )
        # `scales` values should be statically known due to constant folding in shape inference.
        scales = g_op(g, "Concat", scale_1, scale_2, axis_i=0)
        resize = g_op(
            g,
            "Resize",
            input,
            none,
            scales,
            coordinate_transformation_mode_s="align_corners",
            cubic_coeff_a_f=-0.75,
            mode_s="linear",
            nearest_mode_s="floor",
        )
        self.run_test(g, resize.node(), expect_tensor("Float", shape=(4, 32, 128, 128)))

    def test_reduce_prod_with_axes(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_dtype(torch.long).with_sizes([2]))
        reduce_prod = g_op(g, "ReduceProd", input, axes_i=[0])
        self.run_test(g, reduce_prod.node(), expect_tensor("Long", shape=(1,)))

    def test_reduce_prod_without_axes(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_dtype(torch.long).with_sizes([2]))
        reduce_prod = g_op(g, "ReduceProd", input)
        self.run_test(g, reduce_prod.node(), expect_tensor("Long", shape=(1,)))

    def test_proceeding_nodes_use_prim_pack_padded_output_dtype_correctly(self):
        g = self.create_empty_graph()
        input = g.addInput()
        input.setType(input.type().with_dtype(torch.float).with_sizes([4, 16]))
        length = g.addInput()
        length.setType(length.type().with_dtype(torch.long).with_sizes([4]))
        padded, batch_size = g_op(g, "prim::PackPadded", input, length, outputs=2)
        # `prim::PackPadded` only occurs in tracing mode. Hence its outputs inherits
        # shape and data type from traced graph.
        padded.setType(padded.type().with_dtype(torch.float).with_sizes([None, None]))
        batch_size.setType(batch_size.type().with_dtype(torch.long).with_sizes([None]))
        # `Gather` should use the data type of `batch_size` as the data type of its output.
        gather_idx = self.insert_tensor_constant(g, torch.tensor([0], dtype=torch.long))
        gather = g_op(g, "Gather", batch_size, gather_idx, axis_i=0)
        self.run_test(g, gather.node(), expect_tensor("Long", shape=(None,)))

    def test_squeeze_after_dynamic_if(self):
        from torch.onnx.symbolic_opset11 import squeeze as squeeze11

        g = self.create_empty_graph()

        input = g.addInput()
        input.setType(input.type().with_dtype(torch.float).with_sizes([1, None, 5]))

        # Type is intentionally not bool to test that
        # the added "Cast" node doesn't stop shape inference.
        cond = g.addInput()
        cond.setType(input.type().with_dtype(torch.int32).with_sizes([1]))
        if_op, (if_context, else_context), new_node = jit_utils.add_op_with_blocks(
            as_graphcontext(g), "If", cond, n_blocks=2
        )
        block1_output = if_context.op("Add", input, input)
        block2_output = else_context.op("Identity", input)
        utils._add_output_to_block(if_context.block, block1_output)
        utils._add_output_to_block(else_context.block, block2_output)
        if_output = torch._C._jit_pass_fixup_onnx_controlflow_node(
            new_node, _constants.ONNX_TORCHSCRIPT_EXPORTER_MAX_OPSET
        )[0]
        torch._C._jit_pass_onnx_node_shape_type_inference(
            new_node, {}, _constants.ONNX_TORCHSCRIPT_EXPORTER_MAX_OPSET
        )

        # Exporter will add "If" instead of raw "Squeeze" if it does not know
        # that if the dimension it is squeezing has size 1.
        squeezed = squeeze11(as_graphcontext(g), if_output, dim=0)
        assert squeezed.node().kind() == "onnx::Squeeze"
        self.run_test(g, squeezed.node(), expect_tensor("Float", shape=(None, 5)))


class TestONNXCustomOpShapeInference(pytorch_test_common.ExportTestCase):
    def setUp(self):
        super().setUp()
        self.opset_version = _constants.ONNX_TORCHSCRIPT_EXPORTER_MAX_OPSET

    def test_setType_maintains_output_shape_for_single_custom_op(self):
        self.addCleanup(torch.onnx.unregister_custom_op_symbolic, "::linalg_inv", 9)

        class CustomInverse(torch.nn.Module):
            def forward(self, x):
                return torch.inverse(x) + x

        def linalg_inv_settype(g, self):
            return g.op("com.microsoft::Inverse", self).setType(self.type())

        torch.onnx.register_custom_op_symbolic("::linalg_inv", linalg_inv_settype, 9)
        model = CustomInverse()
        x = torch.randn(2, 3, 3)
        f = io.BytesIO()
        torch.onnx.export(
            model,
            (x,),
            f,
            opset_version=self.opset_version,
            custom_opsets={"com.microsoft": 1},
        )

        model_proto = onnx.load(io.BytesIO(f.getvalue()))
        model_value_info = model_proto.graph.value_info
        self.assertIsNotNone(model_value_info)
        assert model_value_info
        dims = model_value_info[0].type.tensor_type.shape.dim
        for i in range(len(dims)):
            # If node output has shape info, it should have dim_value
            # Otherwise, it has dim_params with dynamic shape
            self.assertTrue(dims[i].HasField("dim_value"))
        for dim, rank in zip(dims, x.size()):
            self.assertEqual(dim.dim_value, rank)

    def test_no_setType_for_single_custom_op(self):
        self.addCleanup(torch.onnx.unregister_custom_op_symbolic, "::linalg_inv", 9)

        class CustomInverse(torch.nn.Module):
            def forward(self, x):
                return torch.inverse(x) + x

        def linalg_inv_no_settype(g, self):
            return g.op("com.microsoft::Inverse", self)

        torch.onnx.register_custom_op_symbolic("::linalg_inv", linalg_inv_no_settype, 9)
        model = CustomInverse()
        x = torch.randn(2, 3, 3)
        f = io.BytesIO()
        torch.onnx.export(
            model,
            (x,),
            f,
            opset_version=self.opset_version,
            custom_opsets={"com.microsoft": 1},
        )

        model_proto = onnx.load(io.BytesIO(f.getvalue()))
        model_value_info = model_proto.graph.value_info
        self.assertIsNotNone(model_value_info)
        assert model_value_info
        dims = model_value_info[0].type.tensor_type.shape.dim
        for i in range(len(dims)):
            # If node output has shape info, it should have dim_value
            # Otherwise, it has dim_params with dynamic shape
            self.assertTrue(dims[i].HasField("dim_param"))

    def test_setType_maintains_output_shape_for_single_custom_op_with_dynamic_axes(
        self,
    ):
        self.addCleanup(torch.onnx.unregister_custom_op_symbolic, "::linalg_inv", 9)

        class CustomInverse(torch.nn.Module):
            def forward(self, x):
                return torch.inverse(x) + x

        def linalg_inv_settype(g, self):
            return g.op("com.microsoft::Inverse", self).setType(
                self.type().with_dtype(torch.float).with_sizes([None, 3, 3])
            )

        torch.onnx.register_custom_op_symbolic("::linalg_inv", linalg_inv_settype, 9)
        model = CustomInverse()
        x = torch.randn(2, 3, 3)
        f = io.BytesIO()
        torch.onnx.export(
            model,
            (x,),
            f,
            opset_version=self.opset_version,
            custom_opsets={"com.microsoft": 1},
            input_names=["x"],
            dynamic_axes={"x": {0: "batch"}},
        )

        model_proto = onnx.load(io.BytesIO(f.getvalue()))
        model_value_info = model_proto.graph.value_info
        self.assertIsNotNone(model_value_info)
        assert model_value_info
        dims = model_value_info[0].type.tensor_type.shape.dim
        # The first axe should be dynamic as we defined when exporting
        self.assertTrue(dims[0].HasField("dim_param"))
        for i in range(1, len(dims)):
            # If node output has shape info, it should have dim_value
            # Otherwise, it has dim_params with dynamic shape
            self.assertTrue(dims[i].HasField("dim_value"))
            self.assertEqual(dims[i].dim_value, x.size()[i])

    def test_setType_maintains_output_shape_for_single_custom_op_with_onnx_ops(self):
        self.addCleanup(torch.onnx.unregister_custom_op_symbolic, "::linalg_inv", 9)

        class CustomInverse(torch.nn.Module):
            def forward(self, x, y, z):
                x = torch.inverse(x)
                return x + y + z

        def linalg_inv_settype(g, self):
            return g.op("com.microsoft::Inverse", self).setType(
                self.type().with_dtype(torch.float).with_sizes([2, 3, 10, 10])
            )

        torch.onnx.register_custom_op_symbolic("::linalg_inv", linalg_inv_settype, 9)
        model = CustomInverse()
        x = torch.randn(2, 3, 10, 10)
        y = torch.randn(2, 3, 10, 10)
        z = torch.randn(2, 3, 10, 10)
        f = io.BytesIO()
        torch.onnx.export(
            model,
            (x, y, z),
            f,
            opset_version=self.opset_version,
            custom_opsets={"com.microsoft": 1},
        )

        model_proto = onnx.load(io.BytesIO(f.getvalue()))
        # To validate the shape of inverse Op, we need to find inverse output name,
        # and then use it to identify its value_info for the shape.
        output_name = ""
        for node in model_proto.graph.node:
            if node.op_type == "Inverse":
                output_name = node.output[0]
                break
        assert output_name
        model_value_info = model_proto.graph.value_info
        self.assertIsNotNone(model_value_info)
        assert model_value_info
        for value_info in model_value_info:
            assert value_info.name
            if value_info.name == output_name:
                dims = value_info.type.tensor_type.shape.dim
                for i in range(len(dims)):
                    # If node output has shape info, it should have dim_value
                    # Otherwise, it has dim_params with dynamic shape
                    self.assertTrue(dims[i].HasField("dim_value"))
                for dim, rank in zip(dims, x.size()):
                    self.assertEqual(dim.dim_value, rank)


if __name__ == "__main__":
    common_utils.run_tests()