#!/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.
# ******************************************************************************
"""
CenterNet model definition for detection
"""
__all__ = ["centernet_base", "centernet_voc_pretrained"]
import numpy as np
from keras import Model, initializers
from keras.layers import Conv2D, Add
from cnn2snn import set_akida_version, AkidaVersion
from .. import akidanet18_imagenet
from ..layer_blocks import (separable_conv_block, conv_block, conv_transpose_block,
sepconv_transpose_block)
from ..utils import fetch_file
from ..model_io import load_model, get_model_path
from ..imagenet.imagenet_train import rescale
[docs]def centernet_base(input_shape=(384, 384, 3),
classes=20,
input_scaling=(127, -1),
separable_cutoff=64):
""" A Keras Model implementing the CenterNet architecture, on top of an AkidaNet backbone
Args:
input_shape (tuple, optional): input shape. Defaults to (384, 384, 3).
classes (int, optional): number of output classes. Defaults to 20.
input_scaling (tuple, optional): input scaling. Defaults to (127, -1).
separable_cutoff (int, optional): maximum number of filters for standard Conv layers.
Layers with more filters than this will be defined as separable Convs. Defaults to 64.
Returns:
keras.Model: a Keras Model instance.
"""
def _build_backbone():
base_model = akidanet18_imagenet(input_shape=input_shape,
include_top=False,
pooling="avg",
input_scaling=input_scaling)
# Last three layers immediately prior to stride-2 layers.
# And a final layer, immediately prior to the GAP pooling op
outputs_layer_names = ['stage_0/conv_3/relu', 'stage_1/conv_3/relu',
'pw_stage_2/conv_3/relu', 'pw_stage_3/conv_3/relu']
outputs = [base_model.get_layer(name=name).output for name in outputs_layer_names]
backbone = Model(inputs=base_model.inputs, outputs=outputs)
return backbone
def _get_block_params(n_filt, separable_cutoff):
if n_filt <= separable_cutoff:
neck_convs_ks = (3, 3)
curr_block = conv_block
block_type = "conv"
transpose_block = conv_transpose_block
kwargs = {}
else:
neck_convs_ks = (5, 5)
curr_block = separable_conv_block
block_type = "sepconv"
transpose_block = sepconv_transpose_block
kwargs = {"fused": False}
return neck_convs_ks, curr_block, block_type, transpose_block, kwargs
# This model is only available for Akida 2.0
with set_akida_version(AkidaVersion.v2):
# Create an AkidaNet network without top layers
backbone = _build_backbone()
# Rescale the backbone input shape
if backbone.input.shape[1] != input_shape[0] or backbone.input.shape[2] != input_shape[1]:
backbone = rescale(backbone, [input_shape[0], input_shape[1]])
# Extract skip connections
skips = backbone.output[::-1][1:]
num_deconv_filters = [skip.shape[-1] for skip in skips]
x = backbone.output[-1]
for i, (n_filt) in enumerate(num_deconv_filters):
neck_convs_ks, curr_block, block_type, transpose_block, kwargs = \
_get_block_params(n_filt, separable_cutoff)
x = curr_block(x,
filters=n_filt,
name=f'neck_{block_type}_{i}',
kernel_size=neck_convs_ks,
padding='same',
use_bias=False,
relu_activation='ReLU7.5',
add_batchnorm=True,
**kwargs)
x = transpose_block(x,
filters=n_filt,
name=f"neck_transpose_{block_type}_{i}",
kernel_size=(4, 4),
padding="same",
strides=(2, 2),
use_bias=False,
relu_activation='ReLU7.5',
add_batchnorm=True)
x = Add(name=f"neck_add_{i}")([x, skips[i]])
# Build the head which is composed of 2 consecutive convs
bias_initializer = initializers.Constant(float(-np.log((1 - 0.1) / 0.1)))
init_kernel = initializers.RandomNormal(stddev=0.001, seed=6)
# In the legacy model there is 3 branches of 64 filters each one.
# This could be merged in one branch
x = conv_block(x, 3 * 64, (3, 3),
add_batchnorm=True,
use_bias=False,
relu_activation='ReLU7.5',
padding="same",
name="head_conv_1",
kernel_initializer=init_kernel)
# The output is built by #classes and box coordinates in xywh
x = Conv2D(classes + 4,
(1, 1),
padding="same",
use_bias=True,
name="head_conv_2",
bias_initializer=bias_initializer,
kernel_initializer=init_kernel)(x)
# Build the model
return Model(inputs=backbone.input, outputs=x, name='centernet_base')
[docs]def centernet_voc_pretrained(quantized=True):
"""
Helper method to retrieve an `centernet_base` model that was trained on VOC detection 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.
"""
if quantized:
model_name_v2 = 'centernet_akidanet18_voc_384_i8_w8_a8.h5'
file_hash_v2 = 'cae0f8f36b2d8d01d6049f03185c05d5a9eb3deed47516f7968dc91bb051d81e'
else:
model_name_v2 = 'centernet_akidanet18_voc_384.h5'
file_hash_v2 = '64f220a63aacf653bd672804a5500f98a8dda06029cbe58d327413c9edd4a12f'
model_path, model_name, file_hash = get_model_path("centernet", 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)