openvino/model-optimizer/extensions/middle/FusedBatchNormTraining.py

84 lines
3.8 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.ops.mvn import MVN
from mo.front.common.partial_infer.utils import int64_array
from mo.front.tf.graph_utils import create_op_node_with_second_input
from mo.graph.graph import Graph
from mo.middle.replacement import MiddleReplacementPattern
from mo.ops.const import Const
from mo.ops.reshape import Reshape
from mo.ops.shape import Shape
class FusedBatchNormTraining(MiddleReplacementPattern):
"""
Transformation looks for the BatchNorm layers in training mode and does the following:
1. Fuses batch dimension with one of the spatial dimensions of the input to BatchNorm because batch normalization is
performed over batch dimension also (per channel(features) dimension).
2. Inserts MVN layer.
3. Reshape MVN output back to the original one.
"""
enabled = True
replacement_id = "Fused_Batch_Norm_is_training_true"
force_shape_inference = True
force_clean_up = True
# transformation works for the NHWC layout because transformation inserts Reshape to fuse N and H dimensions
graph_condition = [lambda graph: graph.graph['layout'] == 'NHWC']
def pattern(self):
return dict(
nodes=[
('op', dict(kind='op', op=lambda op: op in ['FusedBatchNorm', 'FusedBatchNormV2', 'FusedBatchNormV3'],
is_training=True))],
edges=[]
)
def replace_pattern(self, graph: Graph, match: dict):
node = match['op']
node.is_training = False
shape = node.in_port(1).data.get_shape()
assert shape is not None, 'The shape of scale input of the BatchNorm node {} is not defined'.format(node.name)
bn_mean = Const(graph, {'name': node.name + '/mean', 'value': np.zeros(shape, dtype=np.float32),
'override_output_shape': True}).create_node()
bn_std = Const(graph, {'name': node.name + '/std', 'value': np.ones(shape, dtype=np.float32),
'override_output_shape': True}).create_node()
node.in_port(3).get_connection().set_source(bn_mean.out_port(0))
node.in_port(4).get_connection().set_source(bn_std.out_port(0))
# save the original shape
original_shape = Shape(graph, {'name': node.in_port(0).get_source().node.soft_get('name')}).create_node()
original_shape.in_port(0).connect(node.in_port(0).get_source())
mvn = MVN(graph, {'name': node.name + '/mvn_', 'eps': node.soft_get('eps', 1e-6),
'override_output_shape': True}).create_node()
node.in_port(0).get_connection().insert_node(mvn)
reshape_4d = create_op_node_with_second_input(graph, Reshape, int64_array([1, -1, 0, 0]),
{'override_output_shape': True,
'name': node.soft_get('name') + '/fused_batch_and_channels'})
mvn.in_port(0).get_connection().insert_node(reshape_4d)
# restore original shape
reshape_back = Reshape(graph, {'name': mvn.soft_get('name') + '/restore_shape',
'override_output_shape': True}).create_node()
reshape_back.in_port(1).connect(original_shape.out_port(0))
mvn.out_port(0).get_connection().insert_node(reshape_back)