#!/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.
# ******************************************************************************
"""
AkidaNet18 model definition for ImageNet classification.
AkidaNet18 architecture is inspired both from AkidaNet and ResNet18: same depth and dimensions than
ResNet18 but without skip connections and using SeparableConvolution layers.
"""
__all__ = ["akidanet18_imagenet", "akidanet18_imagenet_pretrained"]
from keras import Model, regularizers
from keras.layers import Input, Rescaling
from .imagenet_utils import obtain_input_shape
from ..layer_blocks import conv_block, separable_conv_block, dense_block
from ..utils import fetch_file
from ..model_io import load_model, get_model_path
[docs]
def akidanet18_imagenet(input_shape=None,
include_top=True,
pooling=None,
classes=1000,
depths=(4, 4, 4, 4),
dimensions=(64, 128, 256, 512),
input_scaling=(128, -1)):
"""Instantiates the AkidaNet18 architecture.
Note: input preprocessing is included as part of the model (as a Rescaling layer). This model
expects inputs to be float tensors of pixels with values in the [0, 255] range.
Args:
input_shape (tuple, optional): shape tuple. Defaults to None.
include_top (bool, optional): whether to include the fully-connected
layer at the top of the model. Defaults to True.
pooling (str, optional): optional pooling mode for feature extraction
when `include_top` is `False`.
Defaults to None.
* `None` means that the output of the model will be the 4D tensor
output of the last convolutional block.
* `avg` means that global average pooling will be applied to the
output of the last convolutional block, and thus the output of the
model will be a 2D tensor.
classes (int, optional): optional number of classes to classify images
into, only to be specified if `include_top` is `True`. Defaults to 1000.
depth (tuple, optional): number of layers in each stages of the model. The length of the
tuple defines the number of stages. Defaults to (4, 4, 4, 4).
dimensions (tuple, optional): number of filters in each stage on the model. The length of
the tuple must be equal to the length of the `depth` tuple. Defaults to
(64, 128, 256, 512).
input_scaling (tuple, optional): scale factor and offset to apply to
inputs. Defaults to (128, -1). Note that following Akida convention,
the scale factor is an integer used as a divisor.
Returns:
keras.Model: a Keras model for AkidaNet/ImageNet.
Raises:
ValueError: in case of invalid input shape or mismatching `depth` and `dimensions`.
"""
# Sanity checks
stages = len(depths)
if len(dimensions) != stages:
raise ValueError(f"'depth' and 'dimensions' must be of the same length, received: {depths} "
f"and {dimensions}.")
# Define weight regularization, will apply to the convolutional layers and
# to all pointwise weights of separable convolutional layers.
weight_regularizer = regularizers.l2(4e-5)
# Determine proper input shape and default size.
if input_shape is None:
default_size = 224
else:
rows = input_shape[0]
cols = input_shape[1]
if rows == cols and rows in [128, 160, 192, 224]:
default_size = rows
else:
default_size = 224
input_shape = obtain_input_shape(input_shape,
default_size=default_size,
min_size=32,
include_top=include_top)
img_input = Input(shape=input_shape, name="input")
if input_scaling is None:
x = img_input
else:
scale, offset = input_scaling
x = Rescaling(1. / scale, offset, name="rescaling")(img_input)
# ConvNext stem layer: 4x4 kernel with stride 4
x = conv_block(x,
filters=int(dimensions[0]),
name='convnext_stem',
kernel_size=(4, 4),
padding='same',
use_bias=False,
strides=4,
add_batchnorm=True,
relu_activation='ReLU7.5',
kernel_regularizer=weight_regularizer)
# Define the stages
for stage in range(stages):
# Like for AkidaNet, early layers (first 2 stages) are defined as standard Convolutional and
# next layers are SeparableConvolutional layers
if stage < 2:
current_block = conv_block
kwarg = {"kernel_regularizer": weight_regularizer}
else:
current_block = separable_conv_block
kwarg = {"pointwise_regularizer": weight_regularizer, "fused": False}
strides = 2 if stage > 0 else 1
for i in range(depths[stage]):
# First layer in stage comes with strides 2 except in first stage where strides is
# handled by the previous stem
strides = 2 if i == 0 and stage > 0 else 1
# Handle final pooling in last layer of last stage
if stage == stages - 1 and i == depths[stage] - 1:
pool = 'global_avg' if include_top or pooling == 'avg' else None
else:
pool = None
x = current_block(x,
filters=int(dimensions[stage]),
name=f'stage_{stage}/conv_{i}',
kernel_size=(3, 3),
strides=strides,
padding='same',
use_bias=False,
pooling=pool,
add_batchnorm=True,
relu_activation='ReLU7.5',
post_relu_gap=True,
**kwarg)
# Classification layer
if include_top:
x = dense_block(x,
classes,
add_batchnorm=False,
relu_activation=False,
kernel_initializer="he_normal",
name='classifier',
kernel_regularizer=weight_regularizer)
# Create model
return Model(img_input, x, name='akidanet18_%s_%s' % (input_shape[0], classes))
[docs]
def akidanet18_imagenet_pretrained(quantized=True):
"""
Helper method to retrieve an `akidanet18_imagenet` model that was trained on ImageNet dataset.
Args:
quantized (bool, optional): a boolean indicating whether the model should be loaded
quantized or not. Defaults to True.
Returns:
keras.Model: a Keras Model instance.
"""
# Note: cannot be converted to v1 so we should ultimately removed v1 support and only keep v2
if quantized:
model_name_v2 = 'akidanet18_imagenet_224_i8_w8_a8.h5'
file_hash_v2 = '3ee6309b6a8e5aad570fc22d750e9d1e8db82d9a250e9d8f09e009a815db7a3c'
else:
model_name_v2 = 'akidanet18_imagenet_224.h5'
file_hash_v2 = 'e52cf2a20b786a488b75f371b8b89f0c95a383dafc1331eea51b2356552d9704'
model_path, model_name, file_hash = get_model_path("akidanet18", model_name_v2=model_name_v2,
file_hash_v2=file_hash_v2)
model_path = fetch_file(model_path,
fname=model_name,
file_hash=file_hash,
cache_subdir='models')
return load_model(model_path)