def test6(self):
# Original graph
# data(1,64,1)-->Reduce(axis=-2,keep_dims=True, reduce_type=Sum)-->data(1,1,1)
#
# Reference graph
# data(1,61,1)->Reshape(1,1,8,8)->Pool(1,1,1,1)->Reshape(1,1,1)->Power(scale=64)
#
graph = build_graph(nodes_attributes,
[('placeholder_1', 'placeholder_1_data'),
('placeholder_1_data', 'reduce_1'),
('const', 'const_data'),
('const_data', 'reduce_1', {'in': 1}),
('reduce_1', 'reduce_1_data'),
('reduce_1_data', 'concat'),
],
{'placeholder_1': {'shape': int64_array([1, 64, 1])},
'placeholder_1_data': {'shape': int64_array([1, 64, 1])},
'reduce_1': {'keep_dims': True, 'type': 'ReduceSum'},
'const_data': {'value': int64_array([-2])},
'reduce_1_data': {'shape': int64_array([1, 1, 1])},
}, nodes_with_edges_only=True)
graph.graph['layout'] = 'NCHW'
graph_ref = build_graph(nodes_attributes,
[('placeholder_1', 'placeholder_1_data'),
('placeholder_1_data', 'reshape_1'),
('reshape_1_const', 'reshape_1_const_data'),
('reshape_1_const_data', 'reshape_1'),
('reshape_1', 'reshape_1_data'),
('reshape_1_data', 'pooling'),
('pooling', 'pooling_data'),
('pooling_data', 'reshape_2'),
('reshape_2_const', 'reshape_2_const_data'),
('reshape_2_const_data', 'reshape_2'),
('reshape_2', 'reshape_2_data'),
('reshape_2_data', 'power'),
('power', 'power_data'),
('power_data', 'concat'),
],
{'placeholder_1': {'shape': int64_array([1, 64, 1])},
'placeholder_1_data': {'shape': int64_array([1, 64, 1])},
'reshape_1_const': {'value': int64_array([0, 1, 8, 8]), 'shape': int64_array([4])},
'reshape_1_const_data': {'value': int64_array([0, 1, 8, 8]),
'shape': int64_array([4])},
'reshape_1_data': {'shape': int64_array([1, 1, 8, 8])},
'pooling': {'window': int64_array([1, 1, 8, 8])},
'pooling_data': {'shape': int64_array([1, 1, 1, 1])},
'reshape_2_const': {'value': int64_array([0, 1, 1]), 'shape': int64_array([3])},
'reshape_2_const_data': {'value': int64_array([0, 1, 1]), 'shape': int64_array([3])},
'reshape_2_data': {'shape': int64_array([1, 1, 1])},
'power': {'scale': 64.0},
'power_data': {'shape': int64_array([1, 1, 1])},
}, nodes_with_edges_only=True)
ReduceReplacer().find_and_replace_pattern(graph)
shape_inference(graph)
(flag, resp) = compare_graphs(graph, graph_ref, 'concat', check_op_attrs=True)
self.assertTrue(flag, resp)
def test5(self):
# Original graph
# data(1, 16, 64, 64, 64, 4)-->Reduce(axis=[5],keep_dims=False)-->data(1, 16, 64, 64, 64)
#
# Reference graph
# data(1, 16, 64, 64, 64, 4)->Reshape(1*16*64*64, 64, 2, 2)->Pool(1, 1, 2, 2)->Reshape(1, 16, 64, 64, 64)
#
graph = build_graph(nodes_attributes,
[('placeholder_1', 'placeholder_1_data'),
('placeholder_1_data', 'reduce_1'),
('const', 'const_data'),
('const_data', 'reduce_1', {'in': 1}),
('reduce_1', 'reduce_1_data'),
('reduce_1_data', 'concat'),
],
{'placeholder_1': {'shape': int64_array([1, 16, 64, 64, 64, 4])},
'placeholder_1_data': {'shape': int64_array([1, 16, 64, 64, 64, 4])},
'reduce_1': {'keep_dims': False, 'type': 'ReduceMax'},
'const_data': {'value': int64_array([5])},
'reduce_1_data': {'shape': int64_array([1, 16, 64, 64, 64])},
}, nodes_with_edges_only=True)
graph.graph['layout'] = 'NCHW'
graph_ref = build_graph(nodes_attributes,
[('placeholder_1', 'placeholder_1_data'),
('placeholder_1_data', 'reshape_1'),
('reshape_1_const', 'reshape_1_const_data'),
('reshape_1_const_data', 'reshape_1'),
('reshape_1', 'reshape_1_data'),
('reshape_1_data', 'pooling'),
('pooling', 'pooling_data'),
('pooling_data', 'reshape_2'),
('reshape_2_const', 'reshape_2_const_data'),
('reshape_2_const_data', 'reshape_2'),
('reshape_2', 'reshape_2_data'),
('reshape_2_data', 'concat'),
],
{'placeholder_1': {'shape': int64_array([1, 16, 64, 64, 64, 4])},
'placeholder_1_data': {'shape': int64_array([1, 16, 64, 64, 64, 4])},
'reshape_1_const': {'value': int64_array([0, 4194304, 2, 2]),
'shape': int64_array([4])},
'reshape_1_const_data': {'value': int64_array([0, 4194304, 2, 2]),
'shape': int64_array([4])},
'reshape_1_data': {'shape': int64_array([1, 4194304, 2, 2])},
'pooling': {'window': int64_array([1, 1, 2, 2])},
'pooling_data': {'shape': int64_array([1, 4194304, 1, 1])},
'reshape_2_const': {'value': int64_array([0, 16, 64, 64, 64]),
'shape': int64_array([5])},
'reshape_2_const_data': {'value': int64_array([0, 16, 64, 64, 64]),
'shape': int64_array([5])},
'reshape_2_data': {'shape': int64_array([1, 16, 64, 64, 64])},
}, nodes_with_edges_only=True)
ReduceReplacer().find_and_replace_pattern(graph)
shape_inference(graph)
(flag, resp) = compare_graphs(graph, graph_ref, 'concat', check_op_attrs=True)
self.assertTrue(flag, resp)
def test_initial_reshape_dim_normalizer(self, number, expected_output):
window = ReduceReplacer.initial_reshape_dim_normalizer(number)
self.assertIsNotNone(window, "window is None for i={}".format(number))
self.assertEqual(number, np.prod(window), "{} != prod({})".format(number, window))
self.assertEqual(expected_output, window, "{} != {}".format(expected_output, window))