onnx-source-snapshot/onnx/backend/test/case/node/groupnormalization.py

# Copyright (c) ONNX Project Contributors
#
# SPDX-License-Identifier: Apache-2.0
from __future__ import annotations

import numpy as np

import onnx
from onnx.backend.test.case.base import Base
from onnx.backend.test.case.node import expect


# Group normalization's reference implementation
def _group_normalization(x, num_groups, scale, bias, epsilon=1e-5):
    # Assume channel is first dim
    assert x.shape[1] % num_groups == 0
    group_size = x.shape[1] // num_groups
    # Reshape to [N, group_size, C/group_size, H, W, ...]
    new_shape = [x.shape[0], num_groups, group_size, *list(x.shape[2:])]
    x_reshaped = x.reshape(new_shape)
    axes = tuple(range(2, len(new_shape)))
    mean = np.mean(x_reshaped, axis=axes, keepdims=True)
    var = np.var(x_reshaped, axis=axes, keepdims=True)
    x_normalized = ((x_reshaped - mean) / np.sqrt(var + epsilon)).reshape(x.shape)
    dim_ones = (1,) * (len(x.shape) - 2)
    scale = scale.reshape(-1, *dim_ones)
    bias = bias.reshape(-1, *dim_ones)
    return scale * x_normalized + bias


class GroupNormalization(Base):
    @staticmethod
    def export() -> None:
        c = 4
        num_groups = 2
        x = np.random.randn(3, c, 2, 2).astype(np.float32)
        scale = np.random.randn(c).astype(np.float32)
        bias = np.random.randn(c).astype(np.float32)
        y = _group_normalization(x, num_groups, scale, bias).astype(np.float32)

        node = onnx.helper.make_node(
            "GroupNormalization",
            inputs=["x", "scale", "bias"],
            outputs=["y"],
            num_groups=num_groups,
        )

        expect(
            node,
            inputs=[x, scale, bias],
            outputs=[y],
            name="test_group_normalization_example",
        )

    @staticmethod
    def export_epsilon() -> None:
        c = 4
        num_groups = 2
        x = np.random.randn(3, c, 2, 2).astype(np.float32)
        scale = np.random.randn(c).astype(np.float32)
        bias = np.random.randn(c).astype(np.float32)
        epsilon = 1e-2
        y = _group_normalization(x, num_groups, scale, bias, epsilon).astype(np.float32)

        node = onnx.helper.make_node(
            "GroupNormalization",
            inputs=["x", "scale", "bias"],
            outputs=["y"],
            epsilon=epsilon,
            num_groups=num_groups,
        )

        expect(
            node,
            inputs=[x, scale, bias],
            outputs=[y],
            name="test_group_normalization_epsilon",
        )