test/ops/test_batch_norm.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# Copyright 2024 Arm Limited and/or its affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

import logging
import unittest

from typing import Tuple

import torch
from executorch.backends.arm.test import common
from executorch.backends.arm.test.tester.arm_tester import ArmTester
from parameterized import parameterized

logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)

test_data_suite = [
    # (test_name, test_data, [num_features, affine, track_running_stats, weight, bias, running_mean, running_var,] )
    (
        "zeros_affineT_runStatsT_default_weight_bias_mean_var",
        torch.zeros(1, 32, 112, 112),
        [
            32,
            True,
            True,
        ],
    ),
    (
        "zeros_affineF_runStatsT_default_weight_bias_mean_var",
        torch.zeros(1, 32, 112, 112),
        [
            32,
            False,
            True,
        ],
    ),
    (
        "zeros_affineT_runStatsT_rand_weight_bias_mean_var",
        torch.zeros(1, 32, 112, 112),
        [
            32,
            True,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
    (
        "zeros_affineF_runStatsT_rand_weight_bias_mean_var",
        torch.zeros(1, 32, 112, 112),
        [
            32,
            False,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
    (
        "ones_affineT_runStatsT_default_weight_bias_mean_var",
        torch.ones(1, 32, 112, 112),
        [
            32,
            True,
            True,
        ],
    ),
    (
        "ones_affineF_runStatsT_default_weight_bias_mean_var",
        torch.ones(1, 32, 112, 112),
        [
            32,
            False,
            True,
        ],
    ),
    (
        "ones_affineT_runStatsT_rand_weight_bias_mean_var",
        torch.ones(1, 32, 112, 112),
        [
            32,
            True,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
    (
        "ones_affineF_runStatsT_rand_weight_bias_mean_var",
        torch.ones(1, 32, 112, 112),
        [
            32,
            False,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
    (
        "rand_affineT_runStatsT_default_weight_bias_mean_var",
        torch.rand(1, 32, 112, 112),
        [
            32,
            True,
            True,
        ],
    ),
    (
        "rand_affineF_runStatsT_default_weight_bias_mean_var",
        torch.rand(1, 32, 112, 112),
        [
            32,
            False,
            True,
        ],
    ),
    (
        "rand_affineT_runStatsT_rand_weight_bias_mean_var",
        torch.rand(1, 32, 112, 112),
        [
            32,
            True,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
    (
        "rand_affineF_runStatsT_rand_weight_bias_mean_var",
        torch.rand(1, 32, 112, 112),
        [
            32,
            False,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
    (
        "randn_affineT_runStatsT_default_weight_bias_mean_var",
        torch.randn(1, 32, 112, 112),
        [
            32,
            True,
            True,
        ],
    ),
    (
        "randn_affineF_runStatsT_default_weight_bias_mean_var",
        torch.randn(1, 32, 112, 112),
        [
            32,
            False,
            True,
        ],
    ),
    (
        "randn_affineT_runStatsT_rand_weight_bias_mean_var",
        torch.randn(1, 32, 112, 112),
        [
            32,
            True,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
    (
        "randn_affineF_runStatsT_rand_weight_bias_mean_var",
        torch.randn(1, 32, 112, 112),
        [
            32,
            False,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
    # Test some different sizes
    (
        "size_3_4_5_6_affineT_runStatsT_rand_weight_bias_mean_var",
        torch.rand(3, 4, 5, 6),
        [4, True, True, torch.rand(4), torch.rand(4), torch.rand(4), torch.rand(4)],
    ),
    (
        "size_3_4_5_6_affineF_runStatsT_rand_weight_bias_mean_var",
        torch.rand(3, 4, 5, 6),
        [4, True, True, torch.rand(4), torch.rand(4), torch.rand(4), torch.rand(4)],
    ),
    (
        "size_1_3_254_254_affineT_runStatsT_rand_weight_bias_mean_var",
        torch.rand(1, 3, 254, 254),
        [3, True, True, torch.rand(3), torch.rand(3), torch.rand(3), torch.rand(3)],
    ),
    (
        "size_1_3_254_254_affineF_runStatsT_rand_weight_bias_mean_var",
        torch.rand(1, 3, 254, 254),
        [3, True, True, torch.rand(3), torch.rand(3), torch.rand(3), torch.rand(3)],
    ),
    # Test combination of weight and bias
    (
        "check_weight_bias_affineT_runStatsT_none_none",
        torch.rand(1, 32, 112, 112),
        [32, True, True, None, None],
    ),
    (
        "check_weight_bias_affineF_runStatsT_none_none",
        torch.rand(1, 32, 112, 112),
        [32, False, True, None, None],
    ),
    (
        "check_weight_bias_affineT_runStatsT_weight_none",
        torch.rand(1, 32, 112, 112),
        [32, True, True, torch.rand(32)],
    ),
    (
        "check_weight_bias_affineF_runStatsT_weight_none",
        torch.rand(1, 32, 112, 112),
        [32, False, True, torch.rand(32)],
    ),
    (
        "check_weight_bias_affineT_runStatsT_none_bias",
        torch.rand(1, 32, 112, 112),
        [32, True, True, None, torch.rand(32)],
    ),
    (
        "check_weight_bias_affineF_runStatsT_none_bias",
        torch.rand(1, 32, 112, 112),
        [32, False, True, None, torch.rand(32)],
    ),
    (
        "check_weight_bias_affineT_runStatsT_weight_bias",
        torch.rand(1, 32, 112, 112),
        [32, True, True, torch.rand(32), torch.rand(32)],
    ),
    (
        "check_weight_bias_affineF_runStatsT_weight_bias",
        torch.rand(1, 32, 112, 112),
        [32, False, True, torch.rand(32), torch.rand(32)],
    ),
    # Test combination of running_mean and running_var
    (
        "check_mean_var_affineT_runStatsT_none_none",
        torch.randn(1, 32, 112, 112),
        [32, True, True, torch.rand(32), torch.rand(32), None, None],
    ),
    (
        "check_mean_var_affineF_runStatsT_none_none",
        torch.randn(1, 32, 112, 112),
        [32, False, True, torch.rand(32), torch.rand(32), None, None],
    ),
    (
        "check_mean_var_affineT_runStatsT_mean_none",
        torch.randn(1, 32, 112, 112),
        [32, True, True, torch.rand(32), torch.rand(32), torch.rand(32), None],
    ),
    (
        "check_mean_var_affineF_runStatsT_mean_none",
        torch.randn(1, 32, 112, 112),
        [32, False, True, torch.rand(32), torch.rand(32), torch.rand(32), None],
    ),
    (
        "check_mean_var_affineT_runStatsT_none_var",
        torch.randn(1, 32, 112, 112),
        [32, True, True, torch.rand(32), torch.rand(32), None, torch.rand(32)],
    ),
    (
        "check_mean_var_affineF_runStatsT_none_var",
        torch.randn(1, 32, 112, 112),
        [32, False, True, torch.rand(32), torch.rand(32), None, torch.rand(32)],
    ),
    (
        "check_mean_var_affineT_runStatsT_mean_var",
        torch.randn(1, 32, 112, 112),
        [
            32,
            True,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
    (
        "check_mean_var_affineF_runStatsT_mean_var",
        torch.randn(1, 32, 112, 112),
        [
            32,
            False,
            True,
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
            torch.rand(32),
        ],
    ),
]

test_no_stats_data_suite = [
    # (test_name, test_data, [num_features, affine, track_running_stats, weight, bias, running_mean, running_var, ] )
    (
        "zeros_affineT_runStatsF_default_weight_bias",
        torch.zeros(1, 32, 112, 112),
        [
            32,
            True,
            False,
        ],
    ),
    (
        "zeros_affineF_runStatsF_default_weight_bias",
        torch.zeros(1, 32, 112, 112),
        [
            32,
            False,
            False,
        ],
    ),
    (
        "zeros_affineT_runStatsF_rand_weight_bias",
        torch.zeros(1, 32, 112, 112),
        [32, True, False, torch.rand(32), torch.rand(32)],
    ),
    (
        "zeros_affineF_runStatsF_rand_weight_bias",
        torch.zeros(1, 32, 112, 112),
        [32, False, False, torch.rand(32), torch.rand(32)],
    ),
    (
        "ones_affineT_runStatsF_default_weight_bias",
        torch.ones(1, 32, 112, 112),
        [
            32,
            True,
            False,
        ],
    ),
    (
        "ones_affineF_runStatsF_default_weight_bias",
        torch.ones(1, 32, 112, 112),
        [
            32,
            False,
            False,
        ],
    ),
    (
        "ones_affineT_runStatsF_rand_weight_bias",
        torch.ones(1, 32, 112, 112),
        [32, True, False, torch.rand(32), torch.rand(32)],
    ),
    (
        "ones_affineF_runStatsF",
        torch.ones(1, 32, 112, 112),
        [32, False, False, torch.rand(32), torch.rand(32)],
    ),
    (
        "rand_affineT_runStatsF_default_weight_bias",
        torch.rand(1, 32, 112, 112),
        [
            32,
            True,
            False,
        ],
    ),
    (
        "rand_affineF_runStatsF_default_weight_bias",
        torch.rand(1, 32, 112, 112),
        [
            32,
            False,
            False,
        ],
    ),
    (
        "rand_affineT_runStatsF_rand_weight_bias",
        torch.rand(1, 32, 112, 112),
        [32, True, False, torch.rand(32), torch.rand(32)],
    ),
    (
        "rand_affineF_runStatsF_rand_weight_bias",
        torch.rand(1, 32, 112, 112),
        [32, False, False, torch.rand(32), torch.rand(32)],
    ),
    (
        "randn_affineT_runStatsF_default_weight_bias",
        torch.randn(1, 32, 112, 112),
        [
            32,
            True,
            False,
        ],
    ),
    (
        "randn_affineF_runStatsF_default_weight_bias",
        torch.randn(1, 32, 112, 112),
        [
            32,
            False,
            False,
        ],
    ),
    (
        "randn_affineT_runStatsF_rand_weight_bias",
        torch.randn(1, 32, 112, 112),
        [32, True, False, torch.rand(32), torch.rand(32)],
    ),
    (
        "randn_affineF_runStatsF_rand_weight_bias",
        torch.randn(1, 32, 112, 112),
        [32, False, False, torch.rand(32), torch.rand(32)],
    ),
    # Test some different sizes
    (
        "size_3_4_5_6_affineT_runStatsF_rand_weight_bias_mean_var",
        torch.rand(3, 4, 5, 6),
        [4, True, False, torch.rand(4), torch.rand(4)],
    ),
    (
        "size_3_4_5_6_affineF_runStatsF_rand_weight_bias_mean_var",
        torch.rand(3, 4, 5, 6),
        [4, True, False, torch.rand(4), torch.rand(4)],
    ),
    (
        "size_1_3_254_254_affineT_runStatsF_rand_weight_bias_mean_var",
        torch.rand(1, 3, 254, 254),
        [3, True, False, torch.rand(3), torch.rand(3)],
    ),
    (
        "size_1_3_254_254_affineF_runStatsF_rand_weight_bias_mean_var",
        torch.rand(1, 3, 254, 254),
        [3, True, False, torch.rand(3), torch.rand(3)],
    ),
    # Test combination of weight and bias
    (
        "check_weight_bias_affineT_runStatsF_none_none",
        torch.rand(1, 32, 112, 112),
        [32, True, False, None, None],
    ),
    (
        "check_weight_bias_affineF_runStatsF_none_none",
        torch.rand(1, 32, 112, 112),
        [32, False, False, None, None],
    ),
    (
        "check_weight_bias_affineT_runStatsF_weight_none",
        torch.rand(1, 32, 112, 112),
        [32, True, False, torch.rand(32)],
    ),
    (
        "check_weight_bias_affineF_runStatsF_weight_none",
        torch.rand(1, 32, 112, 112),
        [32, False, False, torch.rand(32)],
    ),
    (
        "check_weight_bias_affineT_runStatsF_none_bias",
        torch.rand(1, 32, 112, 112),
        [32, True, False, None, torch.rand(32)],
    ),
    (
        "check_weight_bias_affineF_runStatsF_none_bias",
        torch.rand(1, 32, 112, 112),
        [32, False, False, None, torch.rand(32)],
    ),
    (
        "check_weight_bias_affineT_runStatsF_weight_bias",
        torch.rand(1, 32, 112, 112),
        [32, True, False, torch.rand(32), torch.rand(32)],
    ),
    (
        "check_weight_bias_affineF_runStatsF_weight_bias",
        torch.rand(1, 32, 112, 112),
        [32, False, False, torch.rand(32), torch.rand(32)],
    ),
]


class TestBatchNorm2d(unittest.TestCase):
    """Tests BatchNorm2d."""

    class BatchNorm2d(torch.nn.Module):
        def __init__(
            self,
            num_features: int = 32,
            affine: bool = False,
            track_running_stats: bool = True,
            weights: torch.tensor = None,
            bias: torch.tensor = None,
            running_mean: torch.tensor = None,
            running_var: torch.tensor = None,
        ):
            super().__init__()
            self.batch_norm_2d = torch.nn.BatchNorm2d(
                num_features, affine=affine, track_running_stats=track_running_stats
            )
            if weights is not None:
                self.batch_norm_2d.weight = torch.nn.Parameter(weights)
            if bias is not None:
                self.batch_norm_2d.bias = torch.nn.Parameter(bias)
            if running_mean is not None:
                self.batch_norm_2d.running_mean = running_mean
            if running_var is not None:
                self.batch_norm_2d.running_var = running_var

        def forward(self, x):
            return self.batch_norm_2d(x)

    def _test_batchnorm2d_tosa_MI_pipeline(
        self, module: torch.nn.Module, test_data: Tuple[torch.Tensor]
    ):
        (
            ArmTester(
                module,
                example_inputs=test_data,
                compile_spec=common.get_tosa_compile_spec("TOSA-0.80.0+MI"),
            )
            .export()
            .check_not(["torch.ops.quantized_decomposed"])
            .to_edge()
            .check_count(
                {
                    "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default": 1
                }
            )
            .partition()
            .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
            .check_not(
                [
                    "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default"
                ]
            )
            .to_executorch()
            .run_method_and_compare_outputs(inputs=test_data)
        )

    def _test_batchnorm2d_no_stats_tosa_MI_pipeline(
        self, module: torch.nn.Module, test_data: Tuple[torch.Tensor]
    ):
        (
            ArmTester(
                module,
                example_example_inputs=test_data,
                compile_spec=common.get_tosa_compile_spec("TOSA-0.80.0+MI"),
            )
            .export()
            .check_count({"torch.ops.aten._native_batch_norm_legit.no_stats": 1})
            .check_not(["torch.ops.quantized_decomposed"])
            .to_edge()
            .check_count(
                {
                    "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_stats": 1
                }
            )
            .partition()
            .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
            .check_not(
                [
                    "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_stats"
                ]
            )
            .to_executorch()
            .run_method_and_compare_outputs(inputs=test_data)
        )

    def _test_batchnorm2d_tosa_BI_pipeline(
        self, module: torch.nn.Module, test_data: Tuple[torch.Tensor]
    ):
        (
            ArmTester(
                module,
                example_inputs=test_data,
                compile_spec=common.get_tosa_compile_spec("TOSA-0.80.0+BI"),
            )
            .quantize()
            .export()
            .check_count(
                {"torch.ops.aten._native_batch_norm_legit_no_training.default": 1}
            )
            .check(["torch.ops.quantized_decomposed"])
            .to_edge()
            .check_count(
                {
                    "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default": 1
                }
            )
            .partition()
            .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
            .check_not(
                [
                    "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default"
                ]
            )
            .to_executorch()
            .run_method_and_compare_outputs(inputs=test_data)
        )

    def _test_batchnorm2d_u55_BI_pipeline(
        self, module: torch.nn.Module, test_data: Tuple[torch.Tensor]
    ):
        (
            ArmTester(
                module,
                example_inputs=test_data,
                compile_spec=common.get_u55_compile_spec(),
            )
            .quantize()
            .export()
            .check_count(
                {"torch.ops.aten._native_batch_norm_legit_no_training.default": 1}
            )
            .check(["torch.ops.quantized_decomposed"])
            .to_edge()
            .check_count(
                {
                    "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default": 1
                }
            )
            .partition()
            .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
            .check_not(
                [
                    "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default"
                ]
            )
            .to_executorch()
        )

    @parameterized.expand(test_data_suite)
    def test_batchnorm2d_tosa_MI(
        self,
        test_name: str,
        test_data: torch.Tensor,
        model_params: (
            int
            | Tuple[
                int, bool, bool, torch.tensor, torch.tensor, torch.tensor, torch.tensor
            ]
        ),
    ):
        self._test_batchnorm2d_tosa_MI_pipeline(
            self.BatchNorm2d(*model_params), (test_data,)
        )

    # Expected to fail since not inplemented
    @parameterized.expand(test_no_stats_data_suite)
    @unittest.expectedFailure
    def test_batchnorm2d_no_stats_tosa_MI(
        self,
        test_name: str,
        test_data: torch.Tensor,
        model_params: (
            int
            | Tuple[
                int, bool, bool, torch.tensor, torch.tensor, torch.tensor, torch.tensor
            ]
        ),
    ):
        self._test_batchnorm2d_no_stats_tosa_MI_pipeline(
            self.BatchNorm2d(*model_params), (test_data,)
        )

    # Expected to fail since ArmQuantizer cannot quantize a BatchNorm layer
    # TODO(MLETORCH-100)
    @parameterized.expand(test_data_suite)
    @unittest.skip(
        reason="Expected to fail since ArmQuantizer cannot quantize a BatchNorm layer"
    )
    def test_batchnorm2d_tosa_BI(
        self,
        test_name: str,
        test_data: torch.Tensor,
        model_params: (
            int
            | Tuple[
                int, bool, bool, torch.tensor, torch.tensor, torch.tensor, torch.tensor
            ]
        ),
    ):
        self._test_batchnorm2d_tosa_BI_pipeline(
            self.BatchNorm2d(*model_params), (test_data,)
        )

    # Expected to fail since ArmQuantizer cannot quantize a BatchNorm layer
    # TODO(MLETORCH-100)
    @parameterized.expand(test_data_suite)
    @unittest.skip(
        reason="Expected to fail since ArmQuantizer cannot quantize a BatchNorm layer"
    )
    @unittest.expectedFailure
    def test_batchnorm2d_u55_BI(
        self,
        test_name: str,
        test_data: torch.Tensor,
        model_params: (
            int
            | Tuple[
                int, bool, bool, torch.tensor, torch.tensor, torch.tensor, torch.tensor
            ]
        ),
    ):
        self._test_batchnorm2d_u55_BI_pipeline(
            self.BatchNorm2d(*model_params), (test_data,)
        )