openvino/model-optimizer/unit_tests/extensions/middle/L2NormFusing_test.py

165 lines
8.8 KiB
Python

# Copyright (C) 2018-2021 Intel Corporation
# SPDX-License-Identifier: Apache-2.0
import unittest
import numpy as np
from generator import generator, generate
from extensions.middle.L2NormFusing import L2NormToNorm
from mo.front.common.partial_infer.utils import int64_array
from mo.utils.ir_engine.compare_graphs import compare_graphs
from unit_tests.utils.graph import build_graph_with_attrs
# A list with nodes attributes used to build various graphs.
nodes = [
('l2_normalize_mul', dict(kind='op', op='Mul', name='l2_norm_name')),
('l2_normalize_mul_data', dict(kind='data')),
('maximum', dict(kind='op', op='Maximum')),
('maximum_data', dict(kind='data')),
('maximum_y_const', dict(kind='op', op='Const', value=np.array(12.e-13, dtype=np.float32))),
('maximum_y_data', dict(kind='data', value=np.array(12.e-13, dtype=np.float32))),
('rsqrt_pow', dict(kind='data', value=-0.5)),
('rsqrt', dict(kind='op', op='Pow')),
('rsqrt_data', dict(kind='data')),
('square_pow', dict(kind='op', op='Const', value=2.)),
('square_pow_data', dict(kind='data', value=2.)),
('square', dict(kind='op', op='Pow')),
('sum', dict(kind='op', op='ReduceSum')),
('sum_data', dict(kind='data')),
('sum_axes', dict(kind='op', op='Const')),
# nodes added after replacement
('normalize_node', dict(kind='op', op='NormalizeL2')),
('weights_node', dict(kind='op', op='Const')),
('result', dict(kind='op', op='Result'))
]
edges = [
('input', 'input_data', {'out': 0}),
('input_data', 'square', {'in': 0}),
('square_pow', 'square_pow_data', {'out': 0}),
('square_pow_data', 'square', {'in': 1}),
('square', 'square_data'),
('square_data', 'sum'),
('sum_axes', 'sum_axes_data'),
('sum_axes_data', 'sum'),
('sum', 'sum_data'),
('maximum_y_const', 'maximum_y_data'),
('maximum_y_data', 'maximum'),
('sum_data', 'maximum'),
('maximum', 'maximum_data'),
('maximum_data', 'rsqrt', {'in': 0}),
('rsqrt_pow', 'rsqrt', {'in': 1}),
('rsqrt', 'rsqrt_data'),
('rsqrt_data', 'l2_normalize_mul'),
('input_data', 'l2_normalize_mul'),
('l2_normalize_mul', 'l2_normalize_mul_data'),
('l2_normalize_mul_data', 'result'),
]
edges_after_replacement = [
('input', 'input_data', {'out': 0}),
('input_data', 'normalize_node'),
('weights_node', 'weights_node_data'),
('weights_node_data', 'normalize_node'),
('normalize_node', 'l2_normalize_mul_data'),
('l2_normalize_mul_data', 'result'),
]
@generator
class L2NormToNormTest(unittest.TestCase):
@generate(*[(int64_array([2, 3]), int64_array([1]), 'NCHW'), # NC layout, normalize C dimension
(int64_array([2, 3]), int64_array([1]), 'NHWC'), # NC layout, normalize C dimension
(int64_array([2, 3, 5]), int64_array([1]), 'NCHW'), # NCH layout, normalize C dimension
(int64_array([2, 3, 5]), int64_array([1]), 'NHWC'), # NCH layout, normalize C dimension
(int64_array([2, 3, 5]), int64_array([-1, -2]), 'NHWC'), # NCH layout, normalize CH dimensions
(int64_array([2, 3, 5]), int64_array([-1, -2]), 'NCHW'), # NCH layout, normalize CH dimensions
(int64_array([2, 3, 5]), int64_array([1, 2]), 'NCHW'), # NCH layout, normalize CH dimensions
(int64_array([2, 3, 5]), int64_array([1, 2]), 'NHWC'), # NCH layout, normalize CH dimensions
(int64_array([2, 3, 5, 7]), int64_array([1]), 'NCHW'), # NCHW layout, normalize C dimension
(int64_array([2, 3, 5, 7]), int64_array([-1]), 'NHWC'), # NHWC layout, normalize C dimension
(int64_array([2, 3, 5, 7]), int64_array([3]), 'NHWC'), # NCHW layout, normalize C dimension
(int64_array([2, 3, 5, 7]), int64_array([-1, 1, 2]), 'NCHW'), # NCHW layout, normalize CHW dimensions
(int64_array([2, 3, 5, 7]), int64_array([-3, -2, -1]), 'NHWC'), # NCHW layout, normalize HWC dimensions
])
def test_positive(self, input_shape, axes, layout):
graph = build_graph_with_attrs(nodes + [
('input', dict(kind='op', shape=input_shape, op='Parameter', data_type=np.float32)),
('input_data', dict(kind='data', shape=input_shape, data_type=np.float32)),
('square_data', dict(kind='data', shape=input_shape)),
('sum_axes_data', dict(kind='data', value=axes, shape=None)),
], edges, nodes_with_edges_only=True)
graph.stage = 'middle'
graph.graph['layout'] = layout
L2NormToNorm().find_and_replace_pattern(graph)
graph_ref = build_graph_with_attrs(nodes + [
('input', dict(kind='op', shape=input_shape, op='Parameter', data_type=np.float32)),
('input_data', dict(kind='data', shape=input_shape, data_type=np.float32)),
('weights_node_data', dict(kind='data', value=axes.sort())),
], edges_after_replacement, nodes_with_edges_only=True)
(flag, resp) = compare_graphs(graph, graph_ref, 'result', check_op_attrs=True)
self.assertTrue(graph.node[graph.get_nodes_with_attributes(type='NormalizeL2')[0]]['name'] == 'l2_norm_name')
self.assertTrue(flag, resp)
@generate(*[(int64_array([2]), int64_array([0]), 'NCHW'),
(int64_array([2, 3]), int64_array([0]), 'NCHW'),
(int64_array([2, 3]), int64_array([0]), 'NHWC'),
(int64_array([2, 3]), int64_array([0, 1]), 'NCHW'),
(int64_array([2, 3]), int64_array([0, 1]), 'NHWC'),
(int64_array([2, 3, 5]), int64_array([0]), 'NCHW'),
(int64_array([2, 3, 5]), int64_array([0]), 'NHWC'),
(int64_array([2, 3, 5]), int64_array([-1]), 'NCHW'),
(int64_array([2, 3, 5]), int64_array([-1]), 'NHWC'),
(int64_array([2, 3, 5]), int64_array([0, 1]), 'NCHW'),
(int64_array([2, 3, 5]), int64_array([0, 1]), 'NHWC'),
(int64_array([2, 3, 5]), int64_array([0, 2]), 'NCHW'),
(int64_array([2, 3, 5]), int64_array([0, 2]), 'NHWC'),
(int64_array([2, 3, 5, 7]), int64_array([0]), 'NCHW'),
(int64_array([2, 3, 5, 7]), int64_array([0]), 'NHWC'),
(int64_array([2, 3, 5, 7]), int64_array([2]), 'NCHW'),
(int64_array([2, 3, 5, 7]), int64_array([2]), 'NHWC'),
(int64_array([2, 3, 5, 7]), int64_array([3]), 'NCHW'),
(int64_array([2, 3, 5, 7]), int64_array([1]), 'NHWC'),
(int64_array([2, 3, 5, 7]), int64_array([1, 2]), 'NCHW'),
(int64_array([2, 3, 5, 7]), int64_array([1, -1]), 'NHWC'),
(int64_array([2, 3, 5, 7]), int64_array([1, -1]), 'NCHW'),
(int64_array([2, 3, 5, 7]), int64_array([-2, -1]), 'NHWC'),
(int64_array([2, 3, 5, 7]), int64_array([1, 3]), 'NCHW'),
(int64_array([2, 3, 5, 7]), int64_array([2, 3]), 'NHWC'),
(int64_array([2, 3, 5, 7]), int64_array([0, 1, 2]), 'NCHW'),
(int64_array([2, 3, 5, 7]), int64_array([0, 1, 2]), 'NHWC'),
(int64_array([2, 3, 5, 7]), int64_array([0, 2, 3]), 'NCHW'),
(int64_array([2, 3, 5, 7]), int64_array([0, 2, 3]), 'NHWC'),
(int64_array([2, 3, 5, 7]), int64_array([0, 1, 2, 3]), 'NCHW'),
(int64_array([2, 3, 5, 7]), int64_array([0, 1, 2, 3]), 'NHWC'),
(int64_array([2, 3, 5, 7, 9]), int64_array([1]), 'NCHW'),
(int64_array([2, 3, 5, 7, 9]), int64_array([-1]), 'NHWC'),
(int64_array([2, 3, 5, 7, 9]), int64_array([1, 2, 3, 4]), 'NCHW'),
(int64_array([2, 3, 5, 7, 9]), int64_array([-1, -2, -3, -4]), 'NHWC'),
])
def test_negative(self, input_shape, axes, layout):
graph = build_graph_with_attrs(nodes + [
('input', dict(kind='op', shape=input_shape, op='Parameter', data_type=np.float32)),
('input_data', dict(kind='data', shape=input_shape, data_type=np.float32)),
('square_data', dict(kind='data', shape=input_shape)),
('sum_axes_data', dict(kind='data', value=axes, shape=None)),
], edges, nodes_with_edges_only=True)
graph.stage = 'middle'
graph.graph['layout'] = layout
L2NormToNorm().find_and_replace_pattern(graph)
graph_ref = build_graph_with_attrs(nodes + [
('input', dict(kind='op', shape=input_shape, op='Parameter', data_type=np.float32)),
('input_data', dict(kind='data', shape=input_shape, data_type=np.float32)),
('square_data', dict(kind='data', shape=input_shape)),
('sum_axes_data', dict(kind='data', value=axes, shape=None)),
], edges, nodes_with_edges_only=True)
(flag, resp) = compare_graphs(graph, graph_ref, 'result', check_op_attrs=True)
self.assertTrue(flag, resp)