Source code for akida_models.imagenet.model_mobilenet_edge

#!/usr/bin/env python
# ******************************************************************************
# Copyright 2020 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
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# See the License for the specific language governing permissions and
# limitations under the License.
# ******************************************************************************
This model is an adaptation of the `mobilenet_imagenet` model for edge
applications. It is based on MobileNetBC with top layers replaced by a quantized
spike extractor and a classification layer.

from keras import Model
from keras.layers import Reshape, Activation

from cnn2snn import load_quantized_model, quantize_layer

from ..layer_blocks import separable_conv_block, dense_block
from ..utils import fetch_file


[docs]def mobilenet_edge_imagenet(base_model, classes): """Instantiates a MobileNet-edge architecture. Args: base_model (str/keras.Model): a mobilenet_imagenet quantized model. classes (int): the number of classes for the edge classifier. Returns: keras.Model: a Keras Model instance. """ if isinstance(base_model, str): base_model = load_quantized_model(base_model) try: # Identify the last separable, which is the base model classifier base_classifier = base_model.get_layer("separable_14") # remember the classifier weight bitwidth wq = base_classifier.quantizer.bitwidth except Exception as e: raise ValueError("The base model is not a quantized \ Mobilenet/Imagenet model") from e # Recreate a model with all layers up to the classifier x = base_classifier.input # Add the new end layer with kernel_size (3, 3) instead of (1,1) for # hardware compatibility reasons x = separable_conv_block(x, filters=2048, kernel_size=(3, 3), padding='same', use_bias=False, add_batchnorm=True, name='spike_generator', add_activation=True) # Then add the Akida edge learning layer that will be dropped after x = dense_block(x, classes, name="classification_layer", add_activation=False, add_batchnorm=False, use_bias=False) act_function = 'softmax' if classes > 1 else 'sigmoid' x = Activation(act_function, name=f'act_{act_function}')(x) x = Reshape((classes,), name='reshape_2')(x) # Create model model = Model(inputs=base_model.input, outputs=x, name=f"{}_edge") # Quantize edge layers model = quantize_layer(model, 'spike_generator', wq) model = quantize_layer(model, 'spike_generator/relu', 1) # NOTE: quantization set to 2 here, to be as close as # possible to the Akida native layer that will replace this one, # with binary weights. model = quantize_layer(model, 'classification_layer', 2) return model
[docs]def mobilenet_edge_imagenet_pretrained(): """ Helper method to retrieve a `mobilenet_edge_imagenet` model that was trained on ImageNet dataset. Returns: keras.Model: a Keras Model instance. """ model_name = 'mobilenet_imagenet_224_alpha_50_edge_iq8_wq4_aq4.h5' file_hash = '29ade25ce1907a4dcaa8454a4aff93f3cdc739455ab878156e39e39f12304cbd' model_path = fetch_file(BASE_WEIGHT_PATH + model_name, fname=model_name, file_hash=file_hash, cache_subdir='models') return load_quantized_model(model_path)