def block(inputs):
x = inputs
x = tf.keras.layers.Lambda(lambda a: tf.keras.backend.mean(
a, axis=spatial_dims, keepdims=True))(x)
x = tf.keras.layers.Conv2D(
num_reduced_filters,
kernel_size=[1, 1],
strides=[1, 1],
kernel_initializer=EfficientNetConvInitializer(),
padding="same",
use_bias=True,
)(x)
x = Swish()(x)
# Excite
x = tf.keras.layers.Conv2D(
filters,
kernel_size=[1, 1],
strides=[1, 1],
kernel_initializer=EfficientNetConvInitializer(),
padding="same",
use_bias=True,
)(x)
x = tf.keras.layers.Activation("sigmoid")(x)
out = tf.keras.layers.Multiply()([x, inputs])
return out
def block(inputs):
if expand_ratio != 1:
x = tf.keras.layers.Conv2D(
filters,
kernel_size=[1, 1],
strides=[1, 1],
kernel_initializer=EfficientNetConvInitializer(),
padding="same",
use_bias=False,
)(inputs)
x = tf.keras.layers.BatchNormalization(
axis=channel_axis,
momentum=batch_norm_momentum,
epsilon=batch_norm_epsilon,
)(x)
x = Swish()(x)
else:
x = inputs
x = tf.keras.layers.DepthwiseConv2D(
[kernel_size, kernel_size],
strides=strides,
depthwise_initializer=EfficientNetConvInitializer(),
padding="same",
use_bias=False,
)(x)
x = tf.keras.layers.BatchNormalization(axis=channel_axis,
momentum=batch_norm_momentum,
epsilon=batch_norm_epsilon)(x)
x = Swish()(x)
if has_se:
x = SEBlock(input_filters, se_ratio, expand_ratio, data_format)(x)
# output phase
x = tf.keras.layers.Conv2D(
output_filters,
kernel_size=[1, 1],
strides=[1, 1],
kernel_initializer=EfficientNetConvInitializer(),
padding="same",
use_bias=False,
)(x)
x = tf.keras.layers.BatchNormalization(axis=channel_axis,
momentum=batch_norm_momentum,
epsilon=batch_norm_epsilon)(x)
if id_skip:
if all(s == 1
for s in strides) and (input_filters == output_filters):
# only apply drop_connect if skip presents.
if drop_connect_rate:
x = DropConnect(drop_connect_rate)(x)
x = tf.keras.layers.Add()([x, inputs])
return x
def EfficientNet(input_shape,
block_args_list: List[BlockArgs],
width_coefficient: float,
depth_coefficient: float,
include_top=True,
weights=None,
input_tensor=None,
pooling=None,
classes=1000,
dropout_rate=0.,
drop_connect_rate=0.,
batch_norm_momentum=0.99,
batch_norm_epsilon=1e-3,
depth_divisor=8,
min_depth=None,
data_format=None,
default_size=None,
**kwargs):
"""
Builder model for EfficientNets.
# Arguments:
input_shape: Optional shape tuple, the input shape
depends on the configuration, with a minimum
decided by the number of stride 2 operations.
When None is provided, it defaults to 224.
Considered the "Resolution" parameter from
the paper (inherently Resolution coefficient).
block_args_list: Optional List of BlockArgs, each
of which detail the arguments of the MBConvBlock.
If left as None, it defaults to the blocks
from the paper.
width_coefficient: Determines the number of channels
available per layer. Compound Coefficient that
needs to be found using grid search on a base
configuration model.
depth_coefficient: Determines the number of layers
available to the model. Compound Coefficient that
needs to be found using grid search on a base
configuration model.
include_top: Whether to include the fully-connected
layer at the top of the network.
weights: `None` (random initialization) or
`imagenet` (ImageNet weights)
input_tensor: Optional Keras tensor (i.e. output of
`layers.Input()`)
to use as image input for the model.
pooling: Optional pooling mode for feature extraction
when `include_top` is `False`.
- `None` means that the output of the model
will be the 4D tensor output of the
last convolutional layer.
- `avg` means that global average pooling
will be applied to the output of the
last convolutional layer, and thus
the output of the model will be a
2D tensor.
- `max` means that global max pooling will
be applied.
classes: Optional number of classes to classify images
into, only to be specified if `include_top` is True, and
if no `weights` argument is specified.
dropout_rate: Float, percentage of random dropout.
drop_connect_rate: Float, percentage of random droped
connections.
batch_norm_momentum: Float, default batch normalization
momentum. Obtained from the paper.
batch_norm_epsilon: Float, default batch normalization
epsilon. Obtained from the paper.
depth_divisor: Optional. Used when rounding off the coefficient
scaled channels and depth of the layers.
min_depth: Optional. Minimum depth value in order to
avoid blocks with 0 layers.
data_format: "channels_first" or "channels_last". If left
as None, defaults to the value set in ~/.keras.
default_size: Specifies the default image size of the model
# Raises:
- ValueError: If weights are not in 'imagenet' or None.
- ValueError: If weights are 'imagenet' and `classes` is
not 1000.
# Returns:
A Keras Model.
"""
if not (weights in {'imagenet', None} or os.path.exists(weights)):
raise ValueError('The `weights` argument should be either '
'`None` (random initialization), `imagenet` '
'(pre-training on ImageNet), '
'or the path to the weights file to be loaded.')
if weights == 'imagenet' and include_top and classes != 1000:
raise ValueError('If using `weights` as `"imagenet"` with `include_top` '
'as true, `classes` should be 1000')
if data_format is None:
data_format = K.image_data_format()
if data_format == 'channels_first':
channel_axis = 1
else:
channel_axis = -1
if default_size is None:
default_size = 224
if block_args_list is None:
block_args_list = DEFAULT_BLOCK_LIST
# count number of strides to compute min size
stride_count = 1
for block_args in block_args_list:
if block_args.strides is not None and block_args.strides[0] > 1:
stride_count += 1
min_size = int(2 ** stride_count)
# Determine proper input shape and default size.
input_shape = _obtain_input_shape(input_shape,
default_size=default_size,
min_size=min_size,
data_format=data_format,
require_flatten=include_top,
weights=weights)
# Stem part
if input_tensor is None:
inputs = layers.Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
inputs = layers.Input(tensor=input_tensor, shape=input_shape)
else:
inputs = input_tensor
x = inputs
x = layers.Conv2D(
filters=round_filters(32, width_coefficient,
depth_divisor, min_depth),
kernel_size=[3, 3],
strides=[2, 2],
kernel_initializer=EfficientNetConvInitializer(),
padding='same',
use_bias=False)(x)
x = layers.BatchNormalization(
axis=channel_axis,
momentum=batch_norm_momentum,
epsilon=batch_norm_epsilon)(x)
x = Swish()(x)
num_blocks = sum([block_args.num_repeat for block_args in block_args_list])
drop_connect_rate_per_block = drop_connect_rate / float(num_blocks)
# Blocks part
for block_idx, block_args in enumerate(block_args_list):
assert block_args.num_repeat > 0
# Update block input and output filters based on depth multiplier.
block_args.input_filters = round_filters(block_args.input_filters, width_coefficient, depth_divisor, min_depth)
block_args.output_filters = round_filters(block_args.output_filters, width_coefficient, depth_divisor, min_depth)
block_args.num_repeat = round_repeats(block_args.num_repeat, depth_coefficient)
# The first block needs to take care of stride and filter size increase.
x = MBConvBlock(block_args.input_filters, block_args.output_filters,
block_args.kernel_size, block_args.strides,
block_args.expand_ratio, block_args.se_ratio,
block_args.identity_skip, drop_connect_rate_per_block * block_idx,
batch_norm_momentum, batch_norm_epsilon, data_format)(x)
if block_args.num_repeat > 1:
block_args.input_filters = block_args.output_filters
block_args.strides = [1, 1]
for _ in range(block_args.num_repeat - 1):
x = MBConvBlock(block_args.input_filters, block_args.output_filters,
block_args.kernel_size, block_args.strides,
block_args.expand_ratio, block_args.se_ratio,
block_args.identity_skip, drop_connect_rate_per_block * block_idx,
batch_norm_momentum, batch_norm_epsilon, data_format)(x)
# Head part
x = layers.Conv2D(
filters=round_filters(1280, width_coefficient, depth_coefficient, min_depth),
kernel_size=[1, 1],
strides=[1, 1],
kernel_initializer=EfficientNetConvInitializer(),
padding='same',
use_bias=False)(x)
x = layers.BatchNormalization(
axis=channel_axis,
momentum=batch_norm_momentum,
epsilon=batch_norm_epsilon)(x)
x = Swish()(x)
if include_top:
x = layers.GlobalAveragePooling2D(data_format=data_format)(x)
if dropout_rate > 0:
x = layers.Dropout(dropout_rate)(x)
x = layers.Dense(classes, kernel_initializer=EfficientNetDenseInitializer())(x)
x = layers.Activation('softmax')(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D()(x)
elif pooling == 'max':
x = layers.GlobalMaxPooling2D()(x)
outputs = x
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = get_source_inputs(input_tensor)
model = Model(inputs, outputs)
# Load weights
if weights == 'imagenet':
if default_size == 224:
if include_top:
weights_path = get_file(
'efficientnet-b0.h5',
"https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b0.h5",
cache_subdir='models')
else:
weights_path = get_file(
'efficientnet-b0_notop.h5.h5',
"https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b0_notop.h5",
cache_subdir='models')
model.load_weights(weights_path)
elif default_size == 240:
if include_top:
weights_path = get_file(
'efficientnet-b1.h5',
"https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b1.h5",
cache_subdir='models')
else:
weights_path = get_file(
'efficientnet-b1_notop.h5.h5',
"https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b1_notop.h5",
cache_subdir='models')
model.load_weights(weights_path)
elif default_size == 260:
if include_top:
weights_path = get_file(
'efficientnet-b2.h5',
"https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b2.h5",
cache_subdir='models')
else:
weights_path = get_file(
'efficientnet-b2_notop.h5.h5',
"https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b2_notop.h5",
cache_subdir='models')
model.load_weights(weights_path)
elif default_size == 300:
if include_top:
weights_path = get_file(
'efficientnet-b3.h5',
"https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b3.h5",
cache_subdir='models')
else:
weights_path = get_file(
'efficientnet-b3_notop.h5',
"https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b3_notop.h5",
cache_subdir='models')
model.load_weights(weights_path)
# elif default_size == 380:
# if include_top:
# weights_path = get_file(
# 'efficientnet-b4.h5',
# "https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b4.h5",
# cache_subdir='models')
# else:
# weights_path = get_file(
# 'efficientnet-b4_notoph5',
# "https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b4_notop.h5",
# cache_subdir='models')
# model.load_weights(weights_path)
#
# elif default_size == 456:
# if include_top:
# weights_path = get_file(
# 'efficientnet-b5.h5',
# "https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b5.h5",
# cache_subdir='models')
# else:
# weights_path = get_file(
# 'efficientnet-b5_notop.h5',
# "https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b5_notop.h5",
# cache_subdir='models')
# model.load_weights(weights_path)
#
# elif default_size == 528:
# if include_top:
# weights_path = get_file(
# 'efficientnet-b6.h5',
# "https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b6.h5",
# cache_subdir='models')
# else:
# weights_path = get_file(
# 'efficientnet-b6_notop.h5',
# "https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b6_notop.h5",
# cache_subdir='models')
# model.load_weights(weights_path)
#
# elif default_size == 600:
# if include_top:
# weights_path = get_file(
# 'efficientnet-b7.h5',
# "https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b7.h5",
# cache_subdir='models')
# else:
# weights_path = get_file(
# 'efficientnet-b7_notop.h5',
# "https://github.com/titu1994/keras-efficientnets/releases/download/v0.1/efficientnet-b7_notop.h5",
# cache_subdir='models')
# model.load_weights(weights_path)
else:
raise ValueError('ImageNet weights can only be loaded with EfficientNetB0-7')
elif weights is not None:
model.load_weights(weights)
return model
def EfficientNet(input_shape,
block_args_list: List[BlockArgs],
width_coefficient: float,
depth_coefficient: float,
include_top=True,
weights=None,
input_tensor=None,
pooling=None,
classes=1000,
dropout_rate=0.,
drop_connect_rate=0.,
batch_norm_momentum=0.99,
batch_norm_epsilon=1e-3,
depth_divisor=8,
min_depth=None,
data_format=None,
default_size=None,
**kwargs):
if data_format is None:
data_format = K.image_data_format()
# if data_format == 'channels_first':
# channel_axis = 1
# else:
# channel_axis = -1
if default_size is None:
default_size = 224
if block_args_list is None:
block_args_list = get_default_block_list()
# count number of strides to compute min size
stride_count = 1
for block_args in block_args_list:
if block_args.strides is not None and block_args.strides[0] > 1:
stride_count += 1
min_size = int(2 ** stride_count)
# Determine proper input shape and default size.
input_shape = _obtain_input_shape(input_shape,
default_size=default_size,
min_size=min_size,
data_format=data_format,
require_flatten=include_top,
weights=weights)
# Stem part
if input_tensor is None:
inputs = layers.Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
inputs = layers.Input(tensor=input_tensor, shape=input_shape)
else:
inputs = input_tensor
x = inputs
x = layers.Conv2D(
filters=round_filters(32, width_coefficient,
depth_divisor, min_depth),
kernel_size=[3, 3],
strides=[2, 2],
kernel_initializer=EfficientNetConvInitializer(),
padding='same',
use_bias=False)(x)
x = layers.BatchNormalization(
axis=channel_axis,
momentum=batch_norm_momentum,
epsilon=batch_norm_epsilon)(x)
x = Swish()(x)
num_blocks = sum([block_args.num_repeat for block_args in block_args_list])
drop_connect_rate_per_block = drop_connect_rate / float(num_blocks)
# Blocks part
for block_idx, block_args in enumerate(block_args_list):
assert block_args.num_repeat > 0
# Update block input and output filters based on depth multiplier.
block_args.input_filters = round_filters(block_args.input_filters, width_coefficient, depth_divisor, min_depth)
block_args.output_filters = round_filters(block_args.output_filters, width_coefficient, depth_divisor, min_depth)
block_args.num_repeat = round_repeats(block_args.num_repeat, depth_coefficient)
# The first block needs to take care of stride and filter size increase.
x = MBConvBlock(block_args.input_filters, block_args.output_filters,
block_args.kernel_size, block_args.strides,
block_args.expand_ratio, block_args.se_ratio,
block_args.identity_skip, drop_connect_rate_per_block * block_idx,
batch_norm_momentum, batch_norm_epsilon, data_format)(x)
if block_args.num_repeat > 1:
block_args.input_filters = block_args.output_filters
block_args.strides = [1, 1]
for _ in range(block_args.num_repeat - 1):
x = MBConvBlock(block_args.input_filters, block_args.output_filters,
block_args.kernel_size, block_args.strides,
block_args.expand_ratio, block_args.se_ratio,
block_args.identity_skip, drop_connect_rate_per_block * block_idx,
batch_norm_momentum, batch_norm_epsilon, data_format)(x)
# Head part
x = layers.Conv2D(
filters=round_filters(1280, width_coefficient, depth_coefficient, min_depth),
kernel_size=[1, 1],
strides=[1, 1],
kernel_initializer=EfficientNetConvInitializer(),
padding='same',
use_bias=False)(x)
x = layers.BatchNormalization(
axis=channel_axis,
momentum=batch_norm_momentum,
epsilon=batch_norm_epsilon)(x)
x = Swish()(x)
x = layers.GlobalAveragePooling2D(data_format=data_format)(x)
if dropout_rate > 0:
x = layers.Dropout(dropout_rate)(x)
x = layers.Dense(classes, kernel_initializer=EfficientNetDenseInitializer())(x)
x = layers.Activation('softmax')(x)
outputs = x
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = get_source_inputs(input_tensor)
model = Model(inputs, outputs)
# Load weights
elif weights is not None:
model.load_weights(weights)