970b1301b5
* Removed back phase transformations related to IRv7 * Fixed setting value for the input port using the 'set_value' method * Removed front and middle phase transformations related to IRv7 * Cleanup the rest of the Model Optimizer transformations from IRv7 specific transformations * Final cleanup of the deprecated IR v7 related code * Removed 'blobs_as_input' usage in the Model Optimizer. * Removed function '_fuse_add' from the Model Optimizer since it is not used anymore. * Removed 'keep_in_IR' node attribute for FakeQuantize ops in the MO * Disabled failing gpu_engine.user_context test
321 lines
11 KiB
Python
321 lines
11 KiB
Python
"""
|
|
Copyright (C) 2018-2020 Intel Corporation
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
"""
|
|
|
|
import unittest
|
|
|
|
import numpy as np
|
|
|
|
from extensions.middle.ReluQuantizeFuse import ReluQuantizeFuse, ReluFakeQuantizeMark
|
|
from mo.utils.ir_engine.compare_graphs import compare_graphs
|
|
from mo.utils.unittest.graph import build_graph
|
|
|
|
nodes = {
|
|
# input
|
|
'placeholder': {'type': 'Parameter', 'kind': 'op', 'op': 'Parameter'},
|
|
'placeholder_d': {'value': None, 'shape': None, 'kind': 'data', 'data_type': None},
|
|
|
|
# Relu
|
|
'relu': {'kind': 'op', 'op': 'ReLU'},
|
|
'relu_d': {'value': None, 'shape': None, 'kind': 'data'},
|
|
|
|
# Quantize
|
|
'const_1': {'op': 'Const', 'kind': 'op'},
|
|
'const_1_d': {'kind': 'data', 'value': None},
|
|
'const_2': {'op': 'Const', 'kind': 'op'},
|
|
'const_2_d': {'kind': 'data', 'value': None},
|
|
'const_3': {'op': 'Const', 'kind': 'op'},
|
|
'const_3_d': {'kind': 'data', 'value': None},
|
|
'const_4': {'op': 'Const', 'kind': 'op'},
|
|
'const_4_d': {'kind': 'data', 'value': None},
|
|
|
|
'quantize': {'kind': 'op', 'op': 'FakeQuantize'},
|
|
'quantize_d': {'value': None, 'shape': None, 'kind': 'data'},
|
|
|
|
'quantize_1': {'kind': 'op', 'op': 'FakeQuantize'},
|
|
'quantize_1_d': {'value': None, 'shape': None, 'kind': 'data'},
|
|
|
|
# Result
|
|
'output': {'kind': 'op', 'op': 'Result'},
|
|
'output_1': {'kind': 'op', 'op': 'Result'},
|
|
|
|
# Ops for extra connection expressing
|
|
'extra_op': {'kind': 'op', 'op': 'SomeOp'},
|
|
'extra_data': {'kind': 'data'},
|
|
}
|
|
|
|
i8_edges = [
|
|
# op to data connections
|
|
('placeholder', 'placeholder_d'),
|
|
('relu', 'relu_d', {'out': 0}),
|
|
('const_1', 'const_1_d'),
|
|
('const_2', 'const_2_d'),
|
|
('const_3', 'const_3_d'),
|
|
('const_4', 'const_4_d'),
|
|
('quantize', 'quantize_d'),
|
|
|
|
# data to op connections
|
|
('placeholder_d', 'relu'),
|
|
('relu_d', 'quantize', {'in': 0}),
|
|
('const_1_d', 'quantize', {'in': 1}),
|
|
('const_2_d', 'quantize', {'in': 2}),
|
|
('const_3_d', 'quantize', {'in': 3}),
|
|
('const_4_d', 'quantize', {'in': 4}),
|
|
('quantize_d', 'output'),
|
|
]
|
|
|
|
ref_i8_edges = [
|
|
# op to data connections
|
|
('placeholder', 'placeholder_d'),
|
|
('const_1', 'const_1_d'),
|
|
('const_2', 'const_2_d'),
|
|
('const_3', 'const_3_d'),
|
|
('const_4', 'const_4_d'),
|
|
('quantize', 'quantize_d'),
|
|
|
|
# data to op connections
|
|
('placeholder_d', 'quantize', {'in': 0}),
|
|
('const_1_d', 'quantize', {'in': 1}),
|
|
('const_2_d', 'quantize', {'in': 2}),
|
|
('const_3_d', 'quantize', {'in': 3}),
|
|
('const_4_d', 'quantize', {'in': 4}),
|
|
('quantize_d', 'output'),
|
|
|
|
('placeholder_d', 'relu', {'out': 0}),
|
|
('relu', 'relu_d', {'out': 0}),
|
|
]
|
|
|
|
i1_edges = [
|
|
# op to data connections
|
|
('placeholder', 'placeholder_d'),
|
|
('relu', 'relu_d', {'out': 0}),
|
|
('const_1', 'const_1_d'),
|
|
('const_3', 'const_3_d'),
|
|
('const_4', 'const_4_d'),
|
|
('quantize', 'quantize_d'),
|
|
|
|
# data to op connections
|
|
('placeholder_d', 'relu'),
|
|
('relu_d', 'quantize', {'in': 0}),
|
|
('const_1_d', 'quantize', {'in': 1}),
|
|
('const_1_d', 'quantize', {'in': 2}),
|
|
('const_3_d', 'quantize', {'in': 3}),
|
|
('const_4_d', 'quantize', {'in': 4}),
|
|
('quantize_d', 'output'),
|
|
]
|
|
|
|
ref_i1_edges = [
|
|
# op to data connections
|
|
('placeholder', 'placeholder_d'),
|
|
('const_1', 'const_1_d'),
|
|
('const_3', 'const_3_d'),
|
|
('const_4', 'const_4_d'),
|
|
('quantize', 'quantize_d'),
|
|
|
|
# data to op connections
|
|
('placeholder_d', 'quantize', {'in': 0}),
|
|
('const_1_d', 'quantize', {'in': 1}),
|
|
('const_1_d', 'quantize', {'in': 2}),
|
|
('const_3_d', 'quantize', {'in': 3}),
|
|
('const_4_d', 'quantize', {'in': 4}),
|
|
('quantize_d', 'output'),
|
|
|
|
('placeholder_d', 'relu', {'out': 0}),
|
|
('relu', 'relu_d', {'out': 0}),
|
|
|
|
]
|
|
|
|
relu_extra_output = [
|
|
# op to data connections
|
|
('placeholder', 'placeholder_d'),
|
|
('relu', 'relu_d', {'out': 0}),
|
|
|
|
('const_1', 'const_1_d'),
|
|
('const_3', 'const_3_d'),
|
|
('const_4', 'const_4_d'),
|
|
('quantize', 'quantize_d'),
|
|
|
|
# data to op connections
|
|
('placeholder_d', 'relu'),
|
|
('relu_d', 'quantize', {'in': 0}),
|
|
('const_1_d', 'quantize', {'in': 1}),
|
|
('const_1_d', 'quantize', {'in': 2}),
|
|
('const_3_d', 'quantize', {'in': 3}),
|
|
('const_4_d', 'quantize', {'in': 4}),
|
|
('quantize_d', 'output'),
|
|
|
|
# extra output of relu
|
|
('relu_d', 'extra_op'),
|
|
('extra_op', 'extra_data'),
|
|
('extra_data', 'output_1'),
|
|
]
|
|
|
|
const_extra = [
|
|
# op to data connections
|
|
('placeholder', 'placeholder_d'),
|
|
('relu', 'relu_d', {'out': 0}),
|
|
|
|
('const_1', 'const_1_d'),
|
|
('const_3', 'const_3_d'),
|
|
('const_4', 'const_4_d'),
|
|
('quantize', 'quantize_d'),
|
|
('quantize_1', 'quantize_1_d'),
|
|
|
|
# data to op connections
|
|
('placeholder_d', 'relu', {'out': 0}),
|
|
('relu_d', 'quantize', {'in': 0}),
|
|
('relu_d', 'quantize_1', {'in': 0}),
|
|
|
|
('const_1_d', 'quantize', {'in': 1}),
|
|
('const_1_d', 'quantize', {'in': 2}),
|
|
('const_1_d', 'quantize_1', {'in': 1}),
|
|
('const_1_d', 'quantize_1', {'in': 2}),
|
|
|
|
('const_3_d', 'quantize', {'in': 3}),
|
|
('const_3_d', 'quantize_1', {'in': 3}),
|
|
('const_4_d', 'quantize', {'in': 4}),
|
|
('const_4_d', 'quantize_1', {'in': 4}),
|
|
('quantize_d', 'output'),
|
|
('quantize_1_d', 'output_1'),
|
|
]
|
|
|
|
ref_const_extra = [
|
|
# op to data connections
|
|
('placeholder', 'placeholder_d'),
|
|
('const_1', 'const_1_d'),
|
|
('const_2', 'const_2_d'),
|
|
('const_3', 'const_3_d'),
|
|
('const_4', 'const_4_d'),
|
|
('quantize', 'quantize_d'),
|
|
('quantize_1', 'quantize_1_d'),
|
|
|
|
# data to op connections
|
|
('placeholder_d', 'quantize', {'in': 0, 'out': 0}),
|
|
('placeholder_d', 'quantize_1', {'in': 0, 'out': 0}),
|
|
|
|
('const_1_d', 'quantize', {'out': 0, 'in': 1}),
|
|
('const_1_d', 'quantize', {'out': 0, 'in': 2}),
|
|
('const_2_d', 'quantize_1', {'out': 0, 'in': 1}),
|
|
('const_2_d', 'quantize_1', {'out': 0, 'in': 2}),
|
|
|
|
('const_3_d', 'quantize', {'in': 3}),
|
|
('const_3_d', 'quantize_1', {'in': 3}),
|
|
('const_4_d', 'quantize', {'in': 4}),
|
|
('const_4_d', 'quantize_1', {'in': 4}),
|
|
('quantize_d', 'output'),
|
|
('quantize_1_d', 'output_1'),
|
|
|
|
('placeholder_d', 'relu', {'out': 0}),
|
|
('relu', 'relu_d', {'out': 0}),
|
|
]
|
|
|
|
|
|
class ReluQuantizeFuseTests(unittest.TestCase):
|
|
def test_classic_i8_positive_case(self):
|
|
graph = build_graph(nodes, i8_edges,
|
|
{'const_1_d': {'value': np.zeros([1, 2, 3, 4])}, 'quantize': {'levels': 256}},
|
|
nodes_with_edges_only=True)
|
|
graph.graph['layout'] = 'NHWC'
|
|
graph.stage = 'middle'
|
|
|
|
graph_ref = build_graph(nodes, ref_i8_edges, nodes_with_edges_only=True)
|
|
|
|
ReluFakeQuantizeMark().find_and_replace_pattern(graph)
|
|
ReluQuantizeFuse().find_and_replace_pattern(graph)
|
|
|
|
(flag, resp) = compare_graphs(graph, graph_ref, 'output', check_op_attrs=True)
|
|
self.assertTrue(flag, resp)
|
|
|
|
def test_classic_i8_negative_case(self):
|
|
graph = build_graph(nodes, i8_edges,
|
|
{'const_1_d': {'value': np.full([1, 2, 3, 4], -1)}, 'quantize': {'levels': 256}},
|
|
nodes_with_edges_only=True)
|
|
graph.graph['layout'] = 'NHWC'
|
|
graph.stage = 'middle'
|
|
|
|
graph_ref = build_graph(nodes, i8_edges, nodes_with_edges_only=True)
|
|
|
|
ReluFakeQuantizeMark().find_and_replace_pattern(graph)
|
|
ReluQuantizeFuse().find_and_replace_pattern(graph)
|
|
|
|
(flag, resp) = compare_graphs(graph, graph_ref, 'output', check_op_attrs=True)
|
|
self.assertTrue(flag, resp)
|
|
|
|
def test_classic_i1_positive_case(self):
|
|
graph = build_graph(nodes, i1_edges,
|
|
{'const_1_d': {'value': np.zeros([1, 2, 3, 4], dtype=np.float32)},
|
|
'quantize': {'levels': 2}},
|
|
nodes_with_edges_only=True)
|
|
graph.graph['layout'] = 'NHWC'
|
|
graph.stage = 'middle'
|
|
|
|
graph_ref = build_graph(nodes, ref_i1_edges, nodes_with_edges_only=True)
|
|
|
|
ReluFakeQuantizeMark().find_and_replace_pattern(graph)
|
|
ReluQuantizeFuse().find_and_replace_pattern(graph)
|
|
|
|
(flag, resp) = compare_graphs(graph, graph_ref, 'output', check_op_attrs=True)
|
|
self.assertTrue(flag, resp)
|
|
|
|
def test_classic_i1_negative_case(self):
|
|
graph = build_graph(nodes, i1_edges,
|
|
{'const_1_d': {'value': np.full([1, 2, 3, 4], -1, dtype=np.float32)},
|
|
'quantize': {'levels': 2}},
|
|
nodes_with_edges_only=True)
|
|
graph.graph['layout'] = 'NHWC'
|
|
graph.stage = 'middle'
|
|
|
|
graph_ref = build_graph(nodes, ref_i1_edges, nodes_with_edges_only=True)
|
|
|
|
ReluFakeQuantizeMark().find_and_replace_pattern(graph)
|
|
ReluQuantizeFuse().find_and_replace_pattern(graph)
|
|
|
|
(flag, resp) = compare_graphs(graph, graph_ref, 'output', check_op_attrs=True)
|
|
self.assertTrue(flag, resp)
|
|
np.array_equal(np.full([1, 2, 3, 4], float('-inf'), dtype=np.float32), graph_ref.node['const_1_d']['value'])
|
|
|
|
def test_relu_extra_outputs_i1_case(self):
|
|
graph = build_graph(nodes, relu_extra_output,
|
|
{'const_1_d': {'value': np.full([1, 2, 3, 4], -1, dtype=np.float32)},
|
|
'quantize': {'levels': 2}},
|
|
nodes_with_edges_only=True)
|
|
graph.graph['layout'] = 'NHWC'
|
|
graph.stage = 'middle'
|
|
|
|
graph_ref = build_graph(nodes, relu_extra_output, nodes_with_edges_only=True)
|
|
|
|
ReluFakeQuantizeMark().find_and_replace_pattern(graph)
|
|
ReluQuantizeFuse().find_and_replace_pattern(graph)
|
|
|
|
(flag, resp) = compare_graphs(graph, graph_ref, 'relu', check_op_attrs=True)
|
|
self.assertTrue(flag, resp)
|
|
np.array_equal(np.full([1, 2, 3, 4], float('-inf'), dtype=np.float32), graph_ref.node['const_1_d']['value'])
|
|
|
|
def test_const_extra_outputs_i1_case(self):
|
|
graph = build_graph(nodes, const_extra,
|
|
{'const_1_d': {'value': np.full([1, 2, 3, 4], -1, dtype=np.float32)},
|
|
'quantize': {'levels': 2}, 'quantize_1': {'levels': 2}},
|
|
nodes_with_edges_only=True)
|
|
graph.graph['layout'] = 'NHWC'
|
|
graph.stage = 'middle'
|
|
|
|
graph_ref = build_graph(nodes, ref_const_extra, nodes_with_edges_only=True)
|
|
|
|
ReluFakeQuantizeMark().find_and_replace_pattern(graph)
|
|
ReluQuantizeFuse().find_and_replace_pattern(graph)
|
|
|
|
(flag, resp) = compare_graphs(graph, graph_ref, 'relu', check_op_attrs=True)
|
|
self.assertTrue(flag, resp)
|
|
np.array_equal(np.full([1, 2, 3, 4], float('-inf'), dtype=np.float32), graph_ref.node['const_1_d']['value']) |