Evgeny Lazarev
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
60 lines
2.3 KiB
Python
60 lines
2.3 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 numpy as np
|
|
|
|
from mo.back.replacement import BackReplacementPattern
|
|
from mo.graph.graph import Graph
|
|
|
|
|
|
class PackBinaryWeights(BackReplacementPattern):
|
|
enabled = True
|
|
|
|
@staticmethod
|
|
def pattern():
|
|
return dict(
|
|
nodes=[
|
|
('op', dict(kind='op', type='BinaryConvolution'))],
|
|
edges=[]
|
|
)
|
|
|
|
@staticmethod
|
|
def replace_pattern(graph: Graph, match: dict):
|
|
conv = match['op']
|
|
assert len(conv.in_nodes()) == 2
|
|
initial_shape = conv.in_port(1).data.get_shape()
|
|
assert initial_shape is not None
|
|
weights = conv.in_port(1).data.get_value().flatten()
|
|
weights_rounded = np.round(weights)
|
|
assert np.all(np.isclose(weights, weights_rounded))
|
|
assert len(conv.in_node(1).out_nodes()) == 1
|
|
weights_rounded = np.array(weights_rounded, dtype=np.int32) + 1 # -1 --> 0
|
|
# Reversing element in chunks by 8 elements to pack bits correctly
|
|
# First need to pad data with necessary number of element to make the length dividable by 8
|
|
pad = (-len(weights_rounded)) % 8
|
|
weights_rounded = np.array(np.concatenate((weights_rounded, np.zeros([pad]))), dtype=np.int32)
|
|
assert len(weights_rounded) % 8 == 0
|
|
weights_rounded = weights_rounded.reshape([len(weights_rounded) // 8, 8])
|
|
weights_rounded = np.flip(weights_rounded, axis=1)
|
|
weights_rounded = weights_rounded.flatten()
|
|
packed = np.packbits(weights_rounded)
|
|
conv.in_port(1).data.set_value(packed)
|
|
conv['packed_weights'] = 1
|
|
|
|
conv.in_node(1)['force_shape'] = initial_shape.copy()
|
|
conv.in_node(1)['shape'] = initial_shape.copy()
|
|
conv.in_node(1)['force_type'] = 'U1'
|