#!/usr/bin/env python
# ******************************************************************************
# Copyright 2023 Brainchip Holdings Ltd.
#
# 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.
# ******************************************************************************
"""
Helper that prepares a model for quantization.
"""
__all__ = ['sanitize']
from . import (align_rescaling, invert_batchnorm_pooling, fold_batchnorms, remove_zeropadding2d,
invert_relu_maxpool, replace_lambda, convert_conv_to_dw_conv, remove_reshape,
fold_activations_in_add, split_layers_with_relu, split_concat_layers)
[docs]
def sanitize(model):
""" Sanitize a model preparing it for quantization.
This is a wrapping successive calls to several model transformations which aims at making the
model quantization ready.
Args:
model (keras.Model): the input model
Returns:
keras.Model: the sanitized model
"""
# Replace lambda layers
model = replace_lambda(model)
# Splits layers with 'relu' activation into two separate layers
# one without activation and one with a separate 'relu' activation layer.
model = split_layers_with_relu(model)
# Replace Conv2D layers that behave as DepthwiseConv2D to the latest.
model = convert_conv_to_dw_conv(model)
# Multiple Reshape/Flatten removal transformation
model = remove_reshape(model)
# Align Rescaling (if needed)
model = align_rescaling(model)
# Invert ReLU <-> MaxPool layers so that MaxPool comes first
model = invert_relu_maxpool(model)
# Invert BN <-> Pooling layers and fold BN into their preceding layers
model = invert_batchnorm_pooling(model)
model = fold_batchnorms(model)
# Fold ReLUs with no max_value that comes after an Add layer
model = fold_activations_in_add(model)
# Remove unsupported ZeroPadding2D layers and replace them with 'same' padding convolution when
# possible
model = remove_zeropadding2d(model)
# Split Concatenate layers with more than two inputs into multiple Concatenate layers
# with exactly two inputs when possible
model = split_concat_layers(model)
return model