def MobileNetV1(conn,
model_table='MobileNetV1',
n_classes=1000,
n_channels=3,
width=224,
height=224,
random_flip=None,
random_crop=None,
random_mutation=None,
norm_stds=(255 * 0.229, 255 * 0.224, 255 * 0.225),
offsets=(255 * 0.485, 255 * 0.456, 255 * 0.406),
alpha=1,
depth_multiplier=1):
'''
Generates a deep learning model with the MobileNetV1 architecture.
The implementation is revised based on
https://github.com/keras-team/keras-applications/blob/master/keras_applications/mobilenet.py
Parameters
----------
conn : CAS
Specifies the CAS connection object.
model_table : string or dict or CAS table, optional
Specifies the CAS table to store the deep learning model.
n_classes : int, optional
Specifies the number of classes. If None is assigned, the model will
automatically detect the number of classes based on the training set.
Default: 1000
n_channels : int, optional
Specifies the number of the channels (i.e., depth) of the input layer.
Default: 3
width : int, optional
Specifies the width of the input layer.
Default: 32
height : int, optional
Specifies the height of the input layer.
Default: 32
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
norm_stds : double or iter-of-doubles, optional
Specifies a standard deviation for each channel in the input data.
The final input data is normalized with specified means and standard deviations.
Default: (255*0.229, 255*0.224, 255*0.225)
offsets : double or iter-of-doubles, optional
Specifies an offset for each channel in the input data. The final input
data is set after applying scaling and subtracting the specified offsets.
Default: (255*0.485, 255*0.456, 255*0.406)
alpha : int, optional
Specifies the width multiplier in the MobileNet paper
Default: 1
depth_multiplier : int, optional
Specifies the number of depthwise convolution output channels for each input channel.
Default: 1
Returns
-------
:class:`Model`
References
----------
https://arxiv.org/pdf/1605.07146.pdf
'''
def _conv_block(inputs, filters, alpha, kernel=3, stride=1):
"""
Adds an initial convolution layer (with batch normalization
inputs:
Input tensor
filters:
the dimensionality of the output space
alpha: controls the width of the network.
- If `alpha` < 1.0, proportionally decreases the number
of filters in each layer.
- If `alpha` > 1.0, proportionally increases the number
of filters in each layer.
- If `alpha` = 1, default number of filters from the paper
are used at each layer.
kernel:
specifying the width and height of the 2D convolution window.
strides:
the strides of the convolution
"""
filters = int(filters * alpha)
x = Conv2d(filters,
kernel,
act='identity',
include_bias=False,
stride=stride,
name='conv1')(inputs)
x = BN(name='conv1_bn', act='relu')(x)
return x, filters
def _depthwise_conv_block(inputs,
n_groups,
pointwise_conv_filters,
alpha,
depth_multiplier=1,
stride=1,
block_id=1):
"""Adds a depthwise convolution block.
inputs:
Input tensor
n_groups : int
number of groups
pointwise_conv_filters:
the dimensionality of the output space
alpha: controls the width of the network.
- If `alpha` < 1.0, proportionally decreases the number
of filters in each layer.
- If `alpha` > 1.0, proportionally increases the number
of filters in each layer.
- If `alpha` = 1, default number of filters from the paper
are used at each layer.
depth_multiplier:
The number of depthwise convolution output channels
strides: An integer or tuple/list of 2 integers,
specifying the strides of the convolution
block_id: Integer, a unique identification designating
the block number.
"""
pointwise_conv_filters = int(pointwise_conv_filters * alpha)
x = GroupConv2d(n_groups * depth_multiplier,
n_groups,
3,
stride=stride,
act='identity',
include_bias=False,
name='conv_dw_%d' % block_id)(inputs)
x = BN(name='conv_dw_%d_bn' % block_id, act='relu')(x)
x = Conv2d(pointwise_conv_filters,
1,
act='identity',
include_bias=False,
stride=1,
name='conv_pw_%d' % block_id)(x)
x = BN(name='conv_pw_%d_bn' % block_id, act='relu')(x)
return x, pointwise_conv_filters
parameters = locals()
input_parameters = get_layer_options(input_layer_options, parameters)
inp = Input(**input_parameters, name='data')
# the model down-sampled for 5 times by performing stride=2 convolution on
# conv_dw_1, conv_dw_2, conv_dw_4, conv_dw_6, conv_dw_12
# for each block, we use depthwise convolution with kernel=3 and point-wise convolution to save computation
x, depth = _conv_block(inp, 32, alpha, stride=2)
x, depth = _depthwise_conv_block(x,
depth,
64,
alpha,
depth_multiplier,
block_id=1)
x, depth = _depthwise_conv_block(x,
depth,
128,
alpha,
depth_multiplier,
stride=2,
block_id=2)
x, depth = _depthwise_conv_block(x,
depth,
128,
alpha,
depth_multiplier,
block_id=3)
x, depth = _depthwise_conv_block(x,
depth,
256,
alpha,
depth_multiplier,
stride=2,
block_id=4)
x, depth = _depthwise_conv_block(x,
depth,
256,
alpha,
depth_multiplier,
block_id=5)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
stride=2,
block_id=6)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=7)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=8)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=9)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=10)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=11)
x, depth = _depthwise_conv_block(x,
depth,
1024,
alpha,
depth_multiplier,
stride=2,
block_id=12)
x, depth = _depthwise_conv_block(x,
depth,
1024,
alpha,
depth_multiplier,
block_id=13)
x = GlobalAveragePooling2D(name="Global_avg_pool")(x)
x = OutputLayer(n=n_classes)(x)
model = Model(conn, inp, x, model_table)
model.compile()
return model
def MobileNetV2(conn,
model_table='MobileNetV2',
n_classes=1000,
n_channels=3,
width=224,
height=224,
norm_stds=(255 * 0.229, 255 * 0.224, 255 * 0.225),
offsets=(255 * 0.485, 255 * 0.456, 255 * 0.406),
random_flip=None,
random_crop=None,
random_mutation=None,
alpha=1):
'''
Generates a deep learning model with the MobileNetV2 architecture.
The implementation is revised based on
https://github.com/keras-team/keras-applications/blob/master/keras_applications/mobilenet_v2.py
Parameters
----------
conn : CAS
Specifies the CAS connection object.
model_table : string or dict or CAS table, optional
Specifies the CAS table to store the deep learning model.
n_classes : int, optional
Specifies the number of classes. If None is assigned, the model will
automatically detect the number of classes based on the training set.
Default: 1000
n_channels : int, optional
Specifies the number of the channels (i.e., depth) of the input layer.
Default: 3
width : int, optional
Specifies the width of the input layer.
Default: 224
height : int, optional
Specifies the height of the input layer.
Default: 224
norm_stds : double or iter-of-doubles, optional
Specifies a standard deviation for each channel in the input data.
The final input data is normalized with specified means and standard deviations.
Default: (255 * 0.229, 255 * 0.224, 255 * 0.225)
offsets : double or iter-of-doubles, optional
Specifies an offset for each channel in the input data. The final input
data is set after applying scaling and subtracting the specified offsets.
Default: (255*0.485, 255*0.456, 255*0.406)
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
alpha : int, optional
Specifies the width multiplier in the MobileNet paper
Default: 1
alpha : int, optional
Returns
-------
:class:`Model`
References
----------
https://arxiv.org/abs/1801.04381
'''
def _make_divisible(v, divisor, min_value=None):
# make number of channel divisible
if min_value is None:
min_value = divisor
new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_v < 0.9 * v:
new_v += divisor
return new_v
def _inverted_res_block(inputs, in_channels, expansion, stride, alpha,
filters, block_id):
"""
Inverted Residual Block
Parameters
----------
inputs:
Input tensor
in_channels:
Specifies the number of input tensor's channel
expansion:
expansion factor always applied to the input size.
stride:
the strides of the convolution
alpha:
width multiplier.
filters:
the dimensionality of the output space.
block_id:
block id used for naming layers
"""
pointwise_conv_filters = int(filters * alpha)
pointwise_filters = _make_divisible(pointwise_conv_filters, 8)
x = inputs
prefix = 'block_{}_'.format(block_id)
n_groups = in_channels
if block_id:
# Expand
n_groups = expansion * in_channels
x = Conv2d(expansion * in_channels,
1,
include_bias=False,
act='identity',
name=prefix + 'expand')(x)
x = BN(name=prefix + 'expand_BN', act='identity')(x)
else:
prefix = 'expanded_conv_'
# Depthwise
x = GroupConv2d(n_groups,
n_groups,
3,
stride=stride,
act='identity',
include_bias=False,
name=prefix + 'depthwise')(x)
x = BN(name=prefix + 'depthwise_BN', act='relu')(x)
# Project
x = Conv2d(pointwise_filters,
1,
include_bias=False,
act='identity',
name=prefix + 'project')(x)
x = BN(name=prefix + 'project_BN',
act='identity')(x) # identity activation on narrow tensor
if in_channels == pointwise_filters and stride == 1:
return Res(name=prefix + 'add')([inputs, x]), pointwise_filters
return x, pointwise_filters
parameters = locals()
input_parameters = get_layer_options(input_layer_options, parameters)
inp = Input(**input_parameters, name='data')
# compared with mobilenetv1, v2 introduces inverted residual structure.
# and Non-linearities in narrow layers are removed.
# inverted residual block does three convolutins: first is 1*1 convolution, second is depthwise convolution,
# third is 1*1 convolution but without any non-linearity
first_block_filters = _make_divisible(32 * alpha, 8)
x = Conv2d(first_block_filters,
3,
stride=2,
include_bias=False,
name='Conv1',
act='identity')(inp)
x = BN(name='bn_Conv1', act='relu')(x)
x, n_channels = _inverted_res_block(x,
first_block_filters,
filters=16,
alpha=alpha,
stride=1,
expansion=1,
block_id=0)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=24,
alpha=alpha,
stride=2,
expansion=6,
block_id=1)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=24,
alpha=alpha,
stride=1,
expansion=6,
block_id=2)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=32,
alpha=alpha,
stride=2,
expansion=6,
block_id=3)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=32,
alpha=alpha,
stride=1,
expansion=6,
block_id=4)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=32,
alpha=alpha,
stride=1,
expansion=6,
block_id=5)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=64,
alpha=alpha,
stride=2,
expansion=6,
block_id=6)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=64,
alpha=alpha,
stride=1,
expansion=6,
block_id=7)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=64,
alpha=alpha,
stride=1,
expansion=6,
block_id=8)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=64,
alpha=alpha,
stride=1,
expansion=6,
block_id=9)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=96,
alpha=alpha,
stride=1,
expansion=6,
block_id=10)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=96,
alpha=alpha,
stride=1,
expansion=6,
block_id=11)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=96,
alpha=alpha,
stride=1,
expansion=6,
block_id=12)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=160,
alpha=alpha,
stride=2,
expansion=6,
block_id=13)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=160,
alpha=alpha,
stride=1,
expansion=6,
block_id=14)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=160,
alpha=alpha,
stride=1,
expansion=6,
block_id=15)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=320,
alpha=alpha,
stride=1,
expansion=6,
block_id=16)
# no alpha applied to last conv as stated in the paper:
# if the width multiplier is greater than 1 we increase the number of output channels
if alpha > 1.0:
last_block_filters = _make_divisible(1280 * alpha, 8)
else:
last_block_filters = 1280
x = Conv2d(last_block_filters,
1,
include_bias=False,
name='Conv_1',
act='identity')(x)
x = BN(name='Conv_1_bn', act='relu')(x)
x = GlobalAveragePooling2D(name="Global_avg_pool")(x)
x = OutputLayer(n=n_classes)(x)
model = Model(conn, inp, x, model_table)
model.compile()
return model
def EfficientNet(conn, model_table='EfficientNet', n_classes=100, n_channels=3, width=224, height=224,
width_coefficient=1, depth_coefficient=1, dropout_rate=0.2, drop_connect_rate=0, depth_divisor=8,
activation_fn='relu', blocks_args=_MBConv_BLOCKS_ARGS,
offsets=(255*0.406, 255*0.456, 255*0.485), norm_stds=(255*0.225, 255*0.224, 255*0.229),
random_flip=None, random_crop=None, random_mutation=None):
'''
Generates a deep learning model with the EfficientNet architecture.
The implementation is revised based on
https://github.com/keras-team/keras-applications/blob/master/keras_applications/efficientnet.py
Parameters
----------
conn : CAS
Specifies the CAS connection object.
model_table : string or dict or CAS table, optional
Specifies the CAS table to store the deep learning model.
n_classes : int, optional
Specifies the number of classes. If None is assigned, the model will
automatically detect the number of classes based on the training set.
Default: 1000
n_channels : int, optional
Specifies the number of the channels (i.e., depth) of the input layer.
Default: 3
width : int, optional
Specifies the width of the input layer.
Default: 224
height : int, optional
Specifies the height of the input layer.
Default: 224
width_coefficient: double, optional
Specifies the scale coefficient for network width.
Default: 1.0
depth_coefficient: double, optional
Specifies the scale coefficient for network depth.
Default: 1.0
dropout_rate: double, optional
Specifies the dropout rate before final classifier layer.
Default: 0.2
drop_connect_rate: double, optional
Specifies the dropout rate at skip connections.
Default: 0.0
depth_divisor: integer, optional
Specifies the unit of network width.
Default: 8
activation_fn: string, optional
Specifies the activation function
blocks_args: list of dicts
Specifies parameters to construct blocks for the efficientnet model.
offsets : double or iter-of-doubles, optional
Specifies an offset for each channel in the input data. The final input
data is set after applying scaling and subtracting the specified offsets.
Default: (255*0.406, 255*0.456, 255*0.485)
norm_stds : double or iter-of-doubles, optional
Specifies a standard deviation for each channel in the input data.
The final input data is normalized with specified means and standard deviations.
Default: (255*0.225, 255*0.224, 255*0.229)
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
Returns
-------
:class:`Model`
References
----------
https://arxiv.org/pdf/1905.11946.pdf
'''
def round_filters(filters, width_coefficient, depth_divisor):
'''
round the number of the scaled width, which is for width scaling in efficientnet.
Parameters
----------
filters: integer
Specifies the number of filters.
width_coefficient: double
Specifies the scale coefficient for network width.
depth_divisor: integer
Specifies the unit of network width.
'''
filters *= width_coefficient
new_filters = int(filters + depth_divisor / 2) // depth_divisor * depth_divisor
new_filters = max(depth_divisor, new_filters)
# Make sure that round down does not go down by more than 10%.
if new_filters < 0.9 * filters:
new_filters += depth_divisor
return int(new_filters)
def round_repeats(repeats, depth_coefficient):
'''
round the number of the scaled depth, which is for depth scaling in effcientnet.
Parameters
----------
repeats: integer
Specifies the number of repeats for a block.
depth_coefficient: double
Specifies the scale coefficient for a block.
'''
return int(math.ceil(depth_coefficient * repeats))
def _MBConvBlock(inputs, in_channels, out_channels, ksize, stride, expansion, se_ratio, stage_id, block_id,
noskip=False, activation_fn='relu'):
'''
Inverted Residual Block
Parameters
----------
inputs: input tensor
Speecify input tensor for block.
in_channels: integer
Specifies the number of input tensor's channel.
out_channels: integer
Specifies the number of output tensor's channel
ksize:
Specifies the kernel size of the convolution
stride: integer
Specifies the stride of the convolution
expansion: double
Specifies the expansion factor for the input layer.
se_ratio: double
Specifies the ratio to squeeze the input filters for squeeze-and-excitation block.
stage_id: integer
Specifies stage id for naming layers
block_id:
Specifies block id for naming layers
noskip: bool
Specifies whether the skip connection is used. By default, the skip connection is used.
activation_fn:
Specifies activation function
'''
# mobilenetv2 block is also known as inverted residual block, which consists of three convolutions:
# the first is 1*1 convolution for expansion
# the second is depthwise convolution
# the third is 1*1 convolution without any non-linearity for projection
x = inputs
prefix = 'stage_{}_block_{}'.format(stage_id, block_id)
n_groups = in_channels # for expansion=1, n_groups might be different from pointwise_filters
if expansion > 1:
# For MobileNet V2, expansion>1 when stage>0
n_groups = int(expansion * in_channels) ## update n_groups
x = Conv2d(n_groups, 1, include_bias=False, act='identity',
name=prefix + 'expand')(x)
x = BN(name=prefix + 'expand_BN', act='identity')(x)
# Depthwise convolution
x = GroupConv2d(n_groups, n_groups, ksize, stride=stride, act='identity',
include_bias=False, name=prefix + 'depthwise')(x)
x = BN(name=prefix + 'depthwise_BN', act=activation_fn)(x)
# Squeeze-Excitation
if 0 < se_ratio <= 1:
se_input = x # features to be squeezed
x = GlobalAveragePooling2D(name=prefix + "global_avg_pool")(x)
# Squeeze
channels_se = max(1, int(in_channels * se_ratio))
x = Conv2d(channels_se, 1, include_bias=True, act=activation_fn, name=prefix + 'squeeze')(x)
x = Conv2d(n_groups, 1, include_bias=True, act='sigmoid', name=prefix + 'excitation')(x)
x = Reshape(name=prefix + 'reshape', width=n_groups, height=1, depth=1)(x)
x = Scale(name=prefix + 'scale')([se_input, x]) # x = out*w
# Project
x = Conv2d(out_channels, 1, include_bias=False, act='identity', name=prefix + 'project')(x)
x = BN(name=prefix + 'project_BN', act='identity')(x) # identity activation on narrow tensor
# Prepare output for MBConv block
if in_channels == out_channels and stride == 1 and (not noskip):
# dropout can be added.
return Res(name=prefix + 'add_se_residual')([x, inputs])
else:
return x
parameters = locals()
input_parameters = get_layer_options(input_layer_options, parameters)
inp = Input(**input_parameters, name='data')
# refer to Table 1 "EfficientNet-B0 baseline network" in paper:
# "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks"
stage_id = 0
out_channels = round_filters(32, width_coefficient,
depth_divisor) # multiply with width multiplier: width_coefficient
x = Conv2d(out_channels, 3, stride=2, include_bias=False, name='Conv1', act='identity')(inp)
x = BN(name='bn_Conv1', act=activation_fn)(x)
# Create stages with MBConv blocks from stage 1
in_channels = out_channels # number of input channels for first MBblock
stage_id +=1
total_blocks = float(sum(args[2] for args in blocks_args))
for expansion, out_channels, num_blocks, ksize, stride, se_ratio in blocks_args:
out_channels = round_filters(out_channels, width_coefficient, depth_divisor)
num_blocks = round_repeats(num_blocks, depth_coefficient)
strides = [stride] + [1] * (num_blocks - 1)
for block_id, stride in enumerate(strides):
x = _MBConvBlock(x, in_channels, out_channels, ksize, stride, expansion, se_ratio, stage_id, block_id,activation_fn)
in_channels = out_channels # out_channel
stage_id += 1
last_block_filters = round_filters(1280, width_coefficient, depth_divisor)
x = Conv2d(last_block_filters, 1, include_bias=False, name='Conv_top', act='identity')(x)
x = BN(name='Conv_top_bn', act=activation_fn)(x)
x = GlobalAveragePooling2D(name="Global_avg_pool", dropout=dropout_rate)(x)
x = OutputLayer(n=n_classes)(x)
model = Model(conn, inp, x, model_table)
model.compile()
return model
def test_global_pooling1(self):
global_pooling = GlobalAveragePooling2D()
self.assertTrue(global_pooling.config['width'] == 0)
def _MBConvBlock(inputs, in_channels, out_channels, ksize, stride, expansion, se_ratio, stage_id, block_id,
noskip=False, activation_fn='relu'):
'''
Inverted Residual Block
Parameters
----------
inputs: input tensor
Speecify input tensor for block.
in_channels: integer
Specifies the number of input tensor's channel.
out_channels: integer
Specifies the number of output tensor's channel
ksize:
Specifies the kernel size of the convolution
stride: integer
Specifies the stride of the convolution
expansion: double
Specifies the expansion factor for the input layer.
se_ratio: double
Specifies the ratio to squeeze the input filters for squeeze-and-excitation block.
stage_id: integer
Specifies stage id for naming layers
block_id:
Specifies block id for naming layers
noskip: bool
Specifies whether the skip connection is used. By default, the skip connection is used.
activation_fn:
Specifies activation function
'''
# mobilenetv2 block is also known as inverted residual block, which consists of three convolutions:
# the first is 1*1 convolution for expansion
# the second is depthwise convolution
# the third is 1*1 convolution without any non-linearity for projection
x = inputs
prefix = 'stage_{}_block_{}'.format(stage_id, block_id)
n_groups = in_channels # for expansion=1, n_groups might be different from pointwise_filters
if expansion > 1:
# For MobileNet V2, expansion>1 when stage>0
n_groups = int(expansion * in_channels) ## update n_groups
x = Conv2d(n_groups, 1, include_bias=False, act='identity',
name=prefix + 'expand')(x)
x = BN(name=prefix + 'expand_BN', act='identity')(x)
# Depthwise convolution
x = GroupConv2d(n_groups, n_groups, ksize, stride=stride, act='identity',
include_bias=False, name=prefix + 'depthwise')(x)
x = BN(name=prefix + 'depthwise_BN', act=activation_fn)(x)
# Squeeze-Excitation
if 0 < se_ratio <= 1:
se_input = x # features to be squeezed
x = GlobalAveragePooling2D(name=prefix + "global_avg_pool")(x)
# Squeeze
channels_se = max(1, int(in_channels * se_ratio))
x = Conv2d(channels_se, 1, include_bias=True, act=activation_fn, name=prefix + 'squeeze')(x)
x = Conv2d(n_groups, 1, include_bias=True, act='sigmoid', name=prefix + 'excitation')(x)
x = Reshape(name=prefix + 'reshape', width=n_groups, height=1, depth=1)(x)
x = Scale(name=prefix + 'scale')([se_input, x]) # x = out*w
# Project
x = Conv2d(out_channels, 1, include_bias=False, act='identity', name=prefix + 'project')(x)
x = BN(name=prefix + 'project_BN', act='identity')(x) # identity activation on narrow tensor
# Prepare output for MBConv block
if in_channels == out_channels and stride == 1 and (not noskip):
# dropout can be added.
return Res(name=prefix + 'add_se_residual')([x, inputs])
else:
return x
def DenseNet(conn,
model_table='DenseNet',
n_classes=None,
conv_channel=16,
growth_rate=12,
n_blocks=4,
n_cells=4,
n_channels=3,
width=32,
height=32,
scale=1,
random_flip=None,
random_crop=None,
offsets=(85, 111, 139),
random_mutation=None):
'''
Generates a deep learning model with the DenseNet architecture.
Parameters
----------
conn : CAS
Specifies the connection of the CAS connection.
model_table : string
Specifies the name of CAS table to store the model.
n_classes : int, optional
Specifies the number of classes. If None is assigned, the model will
automatically detect the number of classes based on the training set.
Default: None
conv_channel : int, optional
Specifies the number of filters of the first convolution layer.
Default: 16
growth_rate : int, optional
Specifies the growth rate of convolution layers.
Default: 12
n_blocks : int, optional
Specifies the number of DenseNet blocks.
Default: 4
n_cells : int, optional
Specifies the number of dense connection for each DenseNet block.
Default: 4
n_channels : int, optional
Specifies the number of the channels (i.e., depth) of the input layer.
Default: 3
width : int, optional
Specifies the width of the input layer.
Default: 32
height : int, optional
Specifies the height of the input layer.
Default: 32
scale : double, optional
Specifies a scaling factor to be applied to each pixel intensity values.
Default: 1
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
offsets : double or iter-of-doubles, optional
Specifies an offset for each channel in the input data. The final input
data is set after applying scaling and subtracting the specified offsets.
Default: (85, 111, 139)
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
Returns
-------
:class:`Sequential`
References
----------
https://arxiv.org/pdf/1608.06993.pdf
'''
conn.retrieve('loadactionset',
_messagelevel='error',
actionset='deeplearn')
# get all the parms passed in
parameters = locals()
channel_in = conv_channel # number of channel of transition conv layer
model = Sequential(conn=conn, model_table=model_table)
# get the input parameters
input_parameters = get_layer_options(input_layer_options, parameters)
model.add(InputLayer(**input_parameters))
# Top layers
model.add(
Conv2d(conv_channel,
width=3,
act='identity',
include_bias=False,
stride=1))
for i in range(n_blocks):
model.add(
DenseNetBlock(n_cells=n_cells,
kernel_size=3,
n_filter=growth_rate,
stride=1))
# transition block
channel_in += (growth_rate * n_cells)
model.add(BN(act='relu'))
if i != (n_blocks - 1):
model.add(
Conv2d(channel_in,
width=3,
act='identity',
include_bias=False,
stride=1))
model.add(Pooling(width=2, height=2, pool='mean'))
model.add(GlobalAveragePooling2D())
model.add(OutputLayer(act='softmax', n=n_classes))
return model
def DenseNet121(conn,
model_table='DENSENET121',
n_classes=1000,
conv_channel=64,
growth_rate=32,
n_cells=[6, 12, 24, 16],
n_channels=3,
reduction=0.5,
width=224,
height=224,
scale=1,
random_flip=None,
random_crop=None,
offsets=(103.939, 116.779, 123.68),
random_mutation=None):
'''
Generates a deep learning model with the DenseNet121 architecture.
Parameters
----------
conn : CAS
Specifies the connection of the CAS connection.
model_table : string
Specifies the name of CAS table to store the model.
n_classes : int, optional
Specifies the number of classes. If None is assigned, the model will
automatically detect the number of classes based on the training set.
Default: 1000
conv_channel : int, optional
Specifies the number of filters of the first convolution layer.
Default: 64
growth_rate : int, optional
Specifies the growth rate of convolution layers.
Default: 32
n_cells : int array length=4, optional
Specifies the number of dense connection for each DenseNet block.
Default: [6, 12, 24, 16]
reduction : double, optional
Specifies the factor of transition blocks.
Default: 0.5
n_channels : int, optional
Specifies the number of the channels (i.e., depth) of the input layer.
Default: 3.
width : int, optional
Specifies the width of the input layer.
Default: 224.
height : int, optional
Specifies the height of the input layer.
Default: 224.
scale : double, optional
Specifies a scaling factor to be applied to each pixel intensity values.
Default: 1.
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
offsets : double or iter-of-doubles, optional
Specifies an offset for each channel in the input data. The final input
data is set after applying scaling and subtracting the specified offsets.
Default: (103.939, 116.779, 123.68)
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
Returns
-------
:class:`Sequential`
References
----------
https://arxiv.org/pdf/1608.06993.pdf
'''
conn.retrieve('loadactionset',
_messagelevel='error',
actionset='deeplearn')
# get all the parms passed in
parameters = locals()
n_blocks = len(n_cells)
model = Sequential(conn=conn, model_table=model_table)
# get the input parameters
input_parameters = get_layer_options(input_layer_options, parameters)
model.add(InputLayer(**input_parameters))
# Top layers
model.add(
Conv2d(conv_channel,
width=7,
act='identity',
include_bias=False,
stride=2))
model.add(BN(act='relu'))
src_layer = Pooling(width=3, height=3, stride=2, padding=1, pool='max')
model.add(src_layer)
for i in range(n_blocks):
for _ in range(n_cells[i]):
model.add(BN(act='relu'))
model.add(
Conv2d(n_filters=growth_rate * 4,
width=1,
act='identity',
stride=1,
include_bias=False))
model.add(BN(act='relu'))
src_layer2 = Conv2d(n_filters=growth_rate,
width=3,
act='identity',
stride=1,
include_bias=False)
model.add(src_layer2)
src_layer = Concat(act='identity',
src_layers=[src_layer, src_layer2])
model.add(src_layer)
conv_channel += growth_rate
if i != (n_blocks - 1):
# transition block
conv_channel = int(conv_channel * reduction)
model.add(BN(act='relu'))
model.add(
Conv2d(n_filters=conv_channel,
width=1,
act='identity',
stride=1,
include_bias=False))
src_layer = Pooling(width=2, height=2, stride=2, pool='mean')
model.add(src_layer)
model.add(BN(act='identity'))
# Bottom Layers
model.add(GlobalAveragePooling2D())
model.add(OutputLayer(act='softmax', n=n_classes))
return model
def ShuffleNetV1(conn,
model_table='ShuffleNetV1',
n_classes=1000,
n_channels=3,
width=224,
height=224,
norm_stds=(255 * 0.229, 255 * 0.224, 255 * 0.225),
offsets=(255 * 0.485, 255 * 0.456, 255 * 0.406),
random_flip=None,
random_crop=None,
random_mutation=None,
scale_factor=1.0,
num_shuffle_units=[3, 7, 3],
bottleneck_ratio=0.25,
groups=3,
block_act='identity'):
'''
Generates a deep learning model with the ShuffleNetV1 architecture.
The implementation is revised based on https://github.com/scheckmedia/keras-shufflenet/blob/master/shufflenet.py
Parameters
----------
conn : CAS
Specifies the CAS connection object.
model_table : string or dict or CAS table, optional
Specifies the CAS table to store the deep learning model.
n_classes : int, optional
Specifies the number of classes. If None is assigned, the model will
automatically detect the number of classes based on the training set.
Default: 1000
n_channels : int, optional
Specifies the number of the channels (i.e., depth) of the input layer.
Default: 3
width : int, optional
Specifies the width of the input layer.
Default: 32
height : int, optional
Specifies the height of the input layer.
Default: 32
norm_stds : double or iter-of-doubles, optional
Specifies a standard deviation for each channel in the input data.
The final input data is normalized with specified means and standard deviations.
Default: (255 * 0.229, 255 * 0.224, 255 * 0.225)
offsets : double or iter-of-doubles, optional
Specifies an offset for each channel in the input data. The final input
data is set after applying scaling and subtracting the specified offsets.
Default: (255*0.485, 255*0.456, 255*0.406)
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
scale_factor : double
num_shuffle_units: iter-of-int, optional
number of stages (list length) and the number of shufflenet units in a
stage beginning with stage 2 because stage 1 is fixed
e.g. idx 0 contains 3 + 1 (first shuffle unit in each stage differs) shufflenet units for stage 2
idx 1 contains 7 + 1 Shufflenet Units for stage 3 and
idx 2 contains 3 + 1 Shufflenet Units
Default: [3, 7, 3]
bottleneck_ratio : double
bottleneck ratio implies the ratio of bottleneck channels to output channels.
For example, bottleneck ratio = 1 : 4 means the output feature map is 4 times
the width of the bottleneck feature map.
groups: int
Specifies the number of groups per channel
Default : 3
block_act : str
Specifies the activation function after depth-wise convolution and batch normalization layer
Default : 'identity'
Returns
-------
:class:`Model`
References
----------
https://arxiv.org/pdf/1707.01083
'''
def _block(x, channel_map, bottleneck_ratio, repeat=1, groups=1, stage=1):
"""
creates a bottleneck block
Parameters
----------
x:
Input tensor
channel_map:
list containing the number of output channels for a stage
repeat:
number of repetitions for a shuffle unit with stride 1
groups:
number of groups per channel
bottleneck_ratio:
bottleneck ratio implies the ratio of bottleneck channels to output channels.
stage:
stage number
Returns
-------
"""
x = _shuffle_unit(x,
in_channels=channel_map[stage - 2],
out_channels=channel_map[stage - 1],
strides=2,
groups=groups,
bottleneck_ratio=bottleneck_ratio,
stage=stage,
block=1)
for i in range(1, repeat + 1):
x = _shuffle_unit(x,
in_channels=channel_map[stage - 1],
out_channels=channel_map[stage - 1],
strides=1,
groups=groups,
bottleneck_ratio=bottleneck_ratio,
stage=stage,
block=(i + 1))
return x
def _shuffle_unit(inputs,
in_channels,
out_channels,
groups,
bottleneck_ratio,
strides=2,
stage=1,
block=1):
"""
create a shuffle unit
Parameters
----------
inputs:
Input tensor of with `channels_last` data format
in_channels:
number of input channels
out_channels:
number of output channels
strides:
An integer or tuple/list of 2 integers,
groups:
number of groups per channel
bottleneck_ratio: float
bottleneck ratio implies the ratio of bottleneck channels to output channels.
stage:
stage number
block:
block number
"""
prefix = 'stage%d/block%d' % (stage, block)
# if strides >= 2:
# out_channels -= in_channels
# default: 1/4 of the output channel of a ShuffleNet Unit
bottleneck_channels = int(out_channels * bottleneck_ratio)
groups = (1 if stage == 2 and block == 1 else groups)
# x = _group_conv(inputs, in_channels, out_channels = bottleneck_channels,
# groups = (1 if stage == 2 and block == 1 else groups),
# name = '%s/1x1_gconv_1' % prefix)
x = GroupConv2d(bottleneck_channels,
n_groups=(1 if stage == 2 and block == 1 else groups),
act='identity',
width=1,
height=1,
stride=1,
include_bias=False,
name='%s/1x1_gconv_1' % prefix)(inputs)
x = BN(act='relu', name='%s/bn_gconv_1' % prefix)(x)
x = ChannelShuffle(n_groups=groups,
name='%s/channel_shuffle' % prefix)(x)
# depthwise convolutioin
x = GroupConv2d(x.shape[-1],
n_groups=x.shape[-1],
width=3,
height=3,
include_bias=False,
stride=strides,
act='identity',
name='%s/1x1_dwconv_1' % prefix)(x)
x = BN(act=block_act, name='%s/bn_dwconv_1' % prefix)(x)
out_channels = out_channels if strides == 1 else out_channels - in_channels
x = GroupConv2d(out_channels,
n_groups=groups,
width=1,
height=1,
stride=1,
act='identity',
include_bias=False,
name='%s/1x1_gconv_2' % prefix)(x)
x = BN(act=block_act, name='%s/bn_gconv_2' % prefix)(x)
if strides < 2:
ret = Res(act='relu', name='%s/add' % prefix)([x, inputs])
else:
avg = Pooling(width=3,
height=3,
stride=2,
pool='mean',
name='%s/avg_pool' % prefix)(inputs)
ret = Concat(act='relu', name='%s/concat' % prefix)([x, avg])
return ret
out_dim_stage_two = {1: 144, 2: 200, 3: 240, 4: 272, 8: 384}
try:
import numpy as np
except:
raise DLPyError('Please install numpy to use this architecture.')
exp = np.insert(np.arange(0, len(num_shuffle_units), dtype=np.float32), 0,
0)
out_channels_in_stage = 2**exp
out_channels_in_stage *= out_dim_stage_two[
groups] # calculate output channels for each stage
out_channels_in_stage[0] = 24 # first stage has always 24 output channels
out_channels_in_stage *= scale_factor
out_channels_in_stage = out_channels_in_stage.astype(int)
parameters = locals()
input_parameters = get_layer_options(input_layer_options, parameters)
inp = Input(**input_parameters, name='data')
# create shufflenet architecture
x = Conv2d(out_channels_in_stage[0],
3,
include_bias=False,
stride=2,
act="identity",
name="conv1")(inp)
x = BN(act='relu', name='bn1')(x)
x = Pooling(width=3, height=3, stride=2, name="maxpool1")(x)
# create stages containing shufflenet units beginning at stage 2
for stage in range(0, len(num_shuffle_units)):
repeat = num_shuffle_units[stage]
x = _block(x,
out_channels_in_stage,
repeat=repeat,
bottleneck_ratio=bottleneck_ratio,
groups=groups,
stage=stage + 2)
x = GlobalAveragePooling2D(name="Global_avg_pool")(x)
x = OutputLayer(n=n_classes)(x)
model = Model(conn, inputs=inp, outputs=x, model_table=model_table)
model.compile()
return model
def Darknet_Reference(conn,
model_table='Darknet_Reference',
n_classes=1000,
act='leaky',
n_channels=3,
width=224,
height=224,
scale=1.0 / 255,
random_flip='H',
random_crop='UNIQUE',
random_mutation=None):
'''
Generates a deep learning model with the Darknet_Reference architecture.
The head of the model except the last convolutional layer is same as
the head of Tiny Yolov2. Darknet Reference is pre-trained model for
ImageNet classification.
Parameters
----------
conn : CAS
Specifies the connection of the CAS connection.
model_table : string
Specifies the name of CAS table to store the model.
n_classes : int, optional
Specifies the number of classes. If None is assigned, the model will
automatically detect the number of classes based on the training set.
Default: 1000
act : string
Specifies the type of the activation function for the batch
normalization layers and the final convolution layer.
Default: 'leaky'
n_channels : int, optional
Specifies the number of the channels (i.e., depth) of the input layer.
Default: 3.
width : int, optional
Specifies the width of the input layer.
Default: 224.
height : int, optional
Specifies the height of the input layer.
Default: 224.
scale : double, optional
Specifies a scaling factor to be applied to each pixel intensity values.
Default: 1.0 / 255
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
Default: 'h'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
Default: 'unique'
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
Returns
-------
:class:`Sequential`
'''
conn.retrieve('loadactionset',
_messagelevel='error',
actionset='deeplearn')
# get all the parms passed in
parameters = locals()
model = Sequential(conn=conn, model_table=model_table)
# get the input parameters
input_parameters = get_layer_options(input_layer_options, parameters)
model.add(InputLayer(**input_parameters))
# conv1 224
model.add(Conv2d(16, width=3, act='identity', include_bias=False,
stride=1))
model.add(BN(act=act))
model.add(Pooling(width=2, height=2, stride=2, pool='max'))
# conv2 112
model.add(Conv2d(32, width=3, act='identity', include_bias=False,
stride=1))
model.add(BN(act=act))
model.add(Pooling(width=2, height=2, stride=2, pool='max'))
# conv3 56
model.add(Conv2d(64, width=3, act='identity', include_bias=False,
stride=1))
model.add(BN(act=act))
model.add(Pooling(width=2, height=2, stride=2, pool='max'))
# conv4 28
model.add(
Conv2d(128, width=3, act='identity', include_bias=False, stride=1))
model.add(BN(act=act))
model.add(Pooling(width=2, height=2, stride=2, pool='max'))
# conv5 14
model.add(
Conv2d(256, width=3, act='identity', include_bias=False, stride=1))
model.add(BN(act=act))
model.add(Pooling(width=2, height=2, stride=2, pool='max'))
# conv6 7
model.add(
Conv2d(512, width=3, act='identity', include_bias=False, stride=1))
model.add(BN(act=act))
model.add(Pooling(width=2, height=2, stride=1, pool='max'))
# conv7 7
model.add(
Conv2d(1024, width=3, act='identity', include_bias=False, stride=1))
model.add(BN(act=act))
# conv8 7
model.add(Conv2d(1000, width=1, act=act, include_bias=True, stride=1))
model.add(GlobalAveragePooling2D())
model.add(OutputLayer(act='softmax', n=n_classes))
return model