Files
openvino/model-optimizer/extensions/back/ShuffleChannelPatternOptimization.py
2020-10-09 12:16:12 +03:00

279 lines
12 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 extensions.back.FuseTransposesSequence import FuseTransposesSequence
from extensions.ops.depth_to_space import DepthToSpaceOp
from extensions.ops.shufflechannel import ShuffleChannels
from mo.back.replacement import BackReplacementPattern
from mo.front.common.partial_infer.utils import int64_array
from mo.graph.graph import Graph
class ShuffleChannelPatternOptimization(BackReplacementPattern):
enabled = True
force_clean_up = True
def run_after(self):
return [FuseTransposesSequence]
@staticmethod
def pattern():
return dict(
nodes=[
('t_start_order', {'type': 'Const'}),
('t_start_order_d',
{'value': lambda v: v is not None and np.all(np.array_equal(v, [0, 2, 3, 1]))}),
('t_start', {'type': 'Transpose'}),
('t_start_d', {}),
('reshape_dim', {'type': 'Const'}),
('reshape_dim_d',
{'value': lambda v: v is not None and v.size == 5 and np.all(v[0] == -1)}),
('reshape_start', {'type': 'Reshape'}),
('reshape_start_d', {}),
('t_5d_order', {'type': 'Const'}),
('t_5d_order_d', {'value': lambda v: v is not None and np.all(np.array_equal(v, [0, 1, 2, 4, 3]))}),
('t_5d', {'type': 'Transpose'}),
('t_5d_d', {}),
('reshape_1_dim', {'type': 'Const'}),
('reshape_1_dim_d', {'value': lambda v: v is not None and v.size == 4 and np.all(v[0] == -1)}),
('reshape_end', {'type': 'Reshape'}),
('reshape_end_d', {}),
('t_end_order', {'type': 'Const'}),
('t_end_order_d', {'value': lambda v: v is not None and np.all(np.array_equal(v, [0, 3, 1, 2]))}),
('t_end', {'type': 'Transpose'}),
],
edges=[
('t_start_order', 't_start_order_d'),
('t_start_order_d', 't_start', {'in': 1}),
('t_start', 't_start_d'),
('reshape_dim', 'reshape_dim_d'),
('t_start_d', 'reshape_start', {'in': 0}),
('reshape_dim_d', 'reshape_start', {'in': 1}),
('reshape_start', 'reshape_start_d'),
('t_5d_order', 't_5d_order_d'),
('reshape_start_d', 't_5d', {'in': 0}),
('t_5d_order_d', 't_5d', {'in': 1}),
('t_5d', 't_5d_d'),
('reshape_1_dim', 'reshape_1_dim_d'),
('t_5d_d', 'reshape_end', {'in': 0}),
('reshape_1_dim_d', 'reshape_end', {'in': 1}),
('reshape_end', 'reshape_end_d'),
('t_end_order', 't_end_order_d'),
('reshape_end_d', 't_end', {'in': 0}),
('t_end_order_d', 't_end', {'in': 1}),
],
)
@staticmethod
def feature_dim_splitted(short_shape, long_shape):
return all([short_shape[i] == long_shape[i] for i in range(len(short_shape) - 1)]) and \
short_shape[-1] == long_shape[-1] * long_shape[-2]
@staticmethod
def replace_pattern(graph: Graph, match: dict):
reshape_5d = match['reshape_start']
if not ShuffleChannelPatternOptimization.feature_dim_splitted(
short_shape=reshape_5d.in_port(0).data.get_shape(), long_shape=reshape_5d.out_port(0).data.get_shape()):
return
reshape_4d = match['reshape_end']
if not ShuffleChannelPatternOptimization.feature_dim_splitted(
short_shape=reshape_4d.out_port(0).data.get_shape(), long_shape=reshape_4d.in_port(0).data.get_shape()):
return
start = match['t_start']
end = match['t_end']
new_start = match['reshape_start']
new_end = match['reshape_end']
start_source = start.in_port(0).get_connection().get_source()
end_connection = end.out_port(0).get_connection()
new_end.out_port(0).disconnect()
end_connection.set_source(new_end.out_port(0))
start.in_port(0).disconnect()
new_start.in_port(0).disconnect()
new_start.in_port(0).connect(start_source)
match['reshape_dim']['value'] = int64_array(np.take(new_start.in_port(1).data.get_value(), [0, 3, 4, 1, 2]))
match['reshape_dim'].infer(match['reshape_dim'])
new_start.infer(new_start)
match['t_5d_order']['value'] = int64_array([0, 2, 1, 3, 4])
match['t_5d_order'].infer(match['t_5d_order'])
match['t_5d'].infer(match['t_5d'])
match['reshape_1_dim']['value'] = int64_array(np.take(new_end.in_port(1).data.get_value(), [0, 3, 1, 2]))
match['reshape_1_dim'].infer(match['reshape_1_dim'])
class ShuffleChannelFusion(BackReplacementPattern):
"""
FUSION: Reshape->Transpose->Reshape to ShuffleChannel
We are able to perform the fusion if the pattern satisfies the conditions:
1. Pattern input 4D shape is the same as pattern output 4D shape
2. First Reshape splits channel dimension (1 axis) into two dimensions
3. Transpose permutes only split dimensions
4. Second Reshape pack them back
Fixes original models reshape-ability (Smart reshape)
"""
enabled = True
force_clean_up = True
def run_after(self):
return [FuseTransposesSequence]
@staticmethod
def pattern():
return dict(
nodes=[
('reshape_0_pattern', dict(type='Const')),
('reshape_0_pattern_d', dict(value=lambda v: v is not None and v.size == 5 and np.all(v > 0))),
('reshape_0', dict(type='Reshape')),
('reshape_0_d', dict()),
('order', dict(type='Const')),
('order_d', dict(value=lambda v: v is not None and np.array_equal([0, 2, 1, 3, 4], v))),
('transpose', dict(type='Transpose')),
('transpose_d', {}),
('reshape_1_pattern', dict(type='Const')),
('reshape_1_pattern_d', dict(value=lambda v: v is not None and v.size == 4 and np.all(v > 0))),
('reshape_1', dict(type='Reshape')),
],
edges=[
('reshape_0_pattern', 'reshape_0_pattern_d'),
('reshape_0_pattern_d', 'reshape_0'),
('reshape_0', 'reshape_0_d'),
('reshape_0_d', 'transpose'),
('order', 'order_d'),
('order_d', 'transpose'),
('transpose', 'transpose_d'),
('transpose_d', 'reshape_1'),
('reshape_1_pattern', 'reshape_1_pattern_d'),
('reshape_1_pattern_d', 'reshape_1'),
]
)
@staticmethod
def replace_pattern(graph: Graph, match: dict):
channel_splitting_reshape = match['reshape_0']
channel_concating_reshape = match['reshape_1']
initial_shape = channel_splitting_reshape.in_port(0).data.get_shape()
resulting_shape = channel_concating_reshape.in_port(1).data.get_value()
if not np.array_equal(initial_shape, resulting_shape):
return
channel_splitted_out_shape = channel_splitting_reshape.in_port(1).data.get_value()
if not all([initial_shape[i] == channel_splitted_out_shape[j] for i, j in {0: 0, 2: 3, 3: 4}.items()]):
return
name = channel_concating_reshape.soft_get('name', channel_concating_reshape.id)
group = channel_splitted_out_shape[1]
shuffle_channel = ShuffleChannels(graph, {'name': name, 'group': group}).create_node()
channel_concating_reshape.out_port(0).get_connection().set_source(shuffle_channel.out_port(0))
shuffle_channel.in_port(0).connect(channel_splitting_reshape.in_port(0).get_source())
class DepthToSpaceFusion(BackReplacementPattern):
"""
FUSION: Reshape->Transpose->Reshape to DepthToSpace
We are able to perform the fusion if the pattern satisfies the conditions:
1. Pattern has 6D input and 4D output
2. First Reshape splits channel dimension (1 axis) into three dimensions [new_depth, block_size, block_size]
3. Transpose permutes split dimensions with spatial ones
4. Second Reshape pack block size together with spatial dimension
Fixes original models reshape-ability (Smart reshape)
"""
enabled = True
force_clean_up = True
def run_after(self):
return [FuseTransposesSequence]
@staticmethod
def pattern():
return dict(
nodes=[
('reshape_0_pattern', dict(type='Const')),
('reshape_0_pattern_d', dict(value=lambda v: v is not None and v.size == 6 and np.all(v > 0))),
('reshape_0', dict(type='Reshape')),
('reshape_0_d', dict()),
('order', dict(type='Const')),
('order_d', dict(value=lambda v: v is not None and np.array_equal([0, 1, 4, 2, 5, 3], v))),
('transpose', dict(type='Transpose')),
('transpose_d', {}),
('reshape_1_pattern', dict(type='Const')),
('reshape_1_pattern_d', dict(value=lambda v: v is not None and v.size == 4 and np.all(v > 0))),
('reshape_1', dict(type='Reshape')),
],
edges=[
('reshape_0_pattern', 'reshape_0_pattern_d'),
('reshape_0_pattern_d', 'reshape_0'),
('reshape_0', 'reshape_0_d'),
('reshape_0_d', 'transpose'),
('order', 'order_d'),
('order_d', 'transpose'),
('transpose', 'transpose_d'),
('transpose_d', 'reshape_1'),
('reshape_1_pattern', 'reshape_1_pattern_d'),
('reshape_1_pattern_d', 'reshape_1'),
]
)
@staticmethod
def replace_pattern(graph: Graph, match: dict):
channel_splitting_reshape = match['reshape_0']
channel_concating_reshape = match['reshape_1']
initial_shape = channel_splitting_reshape.in_port(0).data.get_shape()
resulting_shape = channel_concating_reshape.in_port(1).data.get_value()
if initial_shape[0] != resulting_shape[0]:
return
channel_splitted_out_shape = channel_splitting_reshape.in_port(1).data.get_value()
if not all([initial_shape[i] == channel_splitted_out_shape[j] for i, j in {0: 0, 2: 4, 3: 5}.items()]) or \
channel_splitted_out_shape[1] != channel_splitted_out_shape[2]:
return
block_size = channel_splitted_out_shape[2]
expected_output_shape = [initial_shape[0], initial_shape[1] // (block_size * block_size),
initial_shape[2] * block_size, initial_shape[3] * block_size]
if not np.array_equal(expected_output_shape, resulting_shape):
return
name = channel_concating_reshape.soft_get('name', channel_concating_reshape.id)
depth_to_space = DepthToSpaceOp(graph,
{'name': name, 'block_size': block_size, 'mode': 'depth_first'}).create_node()
channel_concating_reshape.out_port(0).get_connection().set_source(depth_to_space.out_port(0))
depth_to_space.in_port(0).connect(channel_splitting_reshape.in_port(0).get_source())