def test_1(self):
#
# NHWC NCHW NHWC
# Input->DATA->Transpose->DATA->Transpose->DATA => Input->DATA
#
graph = build_graph(nodes_attributes, [
('placeholder_1', 'placeholder_1_data'),
('placeholder_1_data', 'permute_1'),
('permute_1', 'permute_1_data'),
('permute_1_data', 'permute_2'),
('permute_2', 'permute_2_data'),
('permute_2_data', 'op_output'),
('const_1', 'const_1_data'),
('const_1_data', 'permute_1', {
'in': 1
}),
('const_2', 'const_2_data'),
('const_2_data', 'permute_2', {
'in': 1
}),
], {
'placeholder_1_data': {
'shape': np.array([1, 227, 227, 3])
},
'const_1_data': {
'value': np.array([0, 3, 1, 2])
},
'permute_1_data': {
'shape': np.array([1, 3, 227, 227])
},
'const_2_data': {
'value': np.array([0, 2, 3, 1])
},
'permute_2_data': {
'shape': np.array([1, 227, 227, 3])
},
},
nodes_with_edges_only=True)
graph.graph['layout'] = 'NHWC'
graph.graph['cmd_params'] = Namespace(keep_shape_ops=False)
graph_ref = build_graph(
nodes_attributes, [('placeholder_1', 'placeholder_1_data'),
('placeholder_1_data', 'op_output')],
{'placeholder_1_data': {
'shape': np.array([1, 227, 227, 3])
}},
nodes_with_edges_only=True)
pattern = FuseTransposesSequence()
pattern.find_and_replace_pattern(graph)
(flag, resp) = compare_graphs(graph,
graph_ref,
'placeholder_1_data',
check_op_attrs=True)
self.assertTrue(flag, resp)
def find_and_replace_pattern(self, graph: Graph):
for start_node in graph.pseudo_topological_sort():
matched_nodes = []
if self.is_node_match_for_optimization(start_node):
next_node = start_node
while self.is_node_match_for_optimization(next_node):
matched_nodes.append(next_node)
next_node[self.OPTIMIZED_NODE_FLAG] = True
next_nodes = get_next_operation(next_node)
if len(next_nodes) > 1:
log.debug('There are two consumers of the node {}. Stop matching sequence.'.format(
next_node.soft_get('name')))
break
next_node = next_nodes[0]
# optimize sequence of three or more Transpose-Reshape nodes
if len(matched_nodes) >= 3:
self.optimize_permute_reshape_sequence(graph, matched_nodes)
# run the RemoveRedundantReshapes to remove dummy (NOP) reshapes. After that we can run Transposes fusing
FuseReshapesSequence().find_and_replace_pattern(graph)
RemoveRedundantReshapes().find_and_replace_pattern(graph)
FuseTransposesSequence().find_and_replace_pattern(graph)
