def _separable_conv_block(ip,
filters,
kernel_size=(3, 3),
strides=(1, 1),
block_id=None):
"""Adds 2 blocks of [relu-separable conv-batchnorm].
Arguments:
ip: Input tensor
filters: Number of output filters per layer
kernel_size: Kernel size of separable convolutions
strides: Strided convolution for downsampling
block_id: String block_id
Returns:
A Keras tensor
"""
channel_dim = 1 if backend.image_data_format() == 'channels_first' else -1
with backend.name_scope('separable_conv_block_%s' % block_id):
x = layers.Activation('relu')(ip)
if strides == (2, 2):
x = layers.ZeroPadding2D(
padding=imagenet_utils.correct_pad(x, kernel_size),
name='separable_conv_1_pad_%s' % block_id)(x)
conv_pad = 'valid'
else:
conv_pad = 'same'
x = layers.SeparableConv2D(
filters,
kernel_size,
strides=strides,
name='separable_conv_1_%s' % block_id,
padding=conv_pad,
use_bias=False,
kernel_initializer='he_normal')(
x)
x = layers.BatchNormalization(
axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='separable_conv_1_bn_%s' % (block_id))(
x)
x = layers.Activation('relu')(x)
x = layers.SeparableConv2D(
filters,
kernel_size,
name='separable_conv_2_%s' % block_id,
padding='same',
use_bias=False,
kernel_initializer='he_normal')(
x)
x = layers.BatchNormalization(
axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='separable_conv_2_bn_%s' % (block_id))(
x)
return x
def _inverted_res_block(inputs, expansion, stride, alpha, filters, block_id):
"""Inverted ResNet block."""
channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1
in_channels = backend.int_shape(inputs)[channel_axis]
pointwise_conv_filters = int(filters * alpha)
pointwise_filters = _make_divisible(pointwise_conv_filters, 8)
x = inputs
prefix = 'block_{}_'.format(block_id)
if block_id:
# Expand
x = layers.Conv2D(expansion * in_channels,
kernel_size=1,
padding='same',
use_bias=False,
activation=None,
name=prefix + 'expand')(x)
x = layers.BatchNormalization(axis=channel_axis,
epsilon=1e-3,
momentum=0.999,
name=prefix + 'expand_BN')(x)
x = layers.ReLU(6., name=prefix + 'expand_relu')(x)
else:
prefix = 'expanded_conv_'
# Depthwise
if stride == 2:
x = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(x, 3),
name=prefix + 'pad')(x)
x = layers.DepthwiseConv2D(kernel_size=3,
strides=stride,
activation=None,
use_bias=False,
padding='same' if stride == 1 else 'valid',
name=prefix + 'depthwise')(x)
x = layers.BatchNormalization(axis=channel_axis,
epsilon=1e-3,
momentum=0.999,
name=prefix + 'depthwise_BN')(x)
x = layers.ReLU(6., name=prefix + 'depthwise_relu')(x)
# Project
x = layers.Conv2D(pointwise_filters,
kernel_size=1,
padding='same',
use_bias=False,
activation=None,
name=prefix + 'project')(x)
x = layers.BatchNormalization(axis=channel_axis,
epsilon=1e-3,
momentum=0.999,
name=prefix + 'project_BN')(x)
if in_channels == pointwise_filters and stride == 1:
return layers.Add(name=prefix + 'add')([inputs, x])
return x
def _inverted_res_block(x, expansion, filters, kernel_size, stride, se_ratio,
activation, block_id):
channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1
shortcut = x
prefix = 'expanded_conv/'
infilters = backend.int_shape(x)[channel_axis]
if block_id:
# Expand
prefix = 'expanded_conv_{}/'.format(block_id)
x = layers.Conv2D(_depth(infilters * expansion),
kernel_size=1,
padding='same',
use_bias=False,
name=prefix + 'expand')(x)
x = layers.BatchNormalization(axis=channel_axis,
epsilon=1e-3,
momentum=0.999,
name=prefix + 'expand/BatchNorm')(x)
x = activation(x)
if stride == 2:
x = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(
x, kernel_size),
name=prefix + 'depthwise/pad')(x)
x = layers.DepthwiseConv2D(kernel_size,
strides=stride,
padding='same' if stride == 1 else 'valid',
use_bias=False,
name=prefix + 'depthwise')(x)
x = layers.BatchNormalization(axis=channel_axis,
epsilon=1e-3,
momentum=0.999,
name=prefix + 'depthwise/BatchNorm')(x)
x = activation(x)
if se_ratio:
x = _se_block(x, _depth(infilters * expansion), se_ratio, prefix)
x = layers.Conv2D(filters,
kernel_size=1,
padding='same',
use_bias=False,
name=prefix + 'project')(x)
x = layers.BatchNormalization(axis=channel_axis,
epsilon=1e-3,
momentum=0.999,
name=prefix + 'project/BatchNorm')(x)
if stride == 1 and infilters == filters:
x = layers.Add(name=prefix + 'Add')([shortcut, x])
return x
def _separable_conv_block(ip,
filters,
kernel_size=(3, 3),
strides=(1, 1),
block_id=None):
channel_dim = -1
with backend.name_scope('separable_conv_block_%s' % block_id):
x = layers.Activation('relu')(ip)
if strides == (2, 2):
x = layers.ZeroPadding2D(
padding=imagenet_utils.correct_pad(x, kernel_size),
name='separable_conv_1_pad_%s' % block_id)(x)
conv_pad = 'valid'
else:
conv_pad = 'same'
x = layers.SeparableConv2D(filters,
kernel_size,
strides=strides,
name='separable_conv_1_%s' % block_id,
padding=conv_pad,
use_bias=False,
kernel_initializer='he_normal')(x)
x = layers.BatchNormalization(axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='separable_conv_1_bn_%s' %
(block_id))(x)
x = layers.Activation('relu')(x)
x = layers.SeparableConv2D(filters,
kernel_size,
name='separable_conv_2_%s' % block_id,
padding='same',
use_bias=False,
kernel_initializer='he_normal')(x)
x = layers.BatchNormalization(axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='separable_conv_2_bn_%s' %
(block_id))(x)
return x
def _reduction_a_cell(ip, p, filters, block_id=None):
"""Adds a Reduction cell for NASNet-A (Fig. 4 in the paper).
Arguments:
ip: Input tensor `x`
p: Input tensor `p`
filters: Number of output filters
block_id: String block_id
Returns:
A Keras tensor
"""
channel_dim = 1 if backend.image_data_format() == 'channels_first' else -1
with backend.name_scope('reduction_A_block_%s' % block_id):
p = _adjust_block(p, ip, filters, block_id)
h = layers.Activation('relu')(ip)
h = layers.Conv2D(filters, (1, 1),
strides=(1, 1),
padding='same',
name='reduction_conv_1_%s' % block_id,
use_bias=False,
kernel_initializer='he_normal')(h)
h = layers.BatchNormalization(axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='reduction_bn_1_%s' % block_id)(h)
h3 = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(h, 3),
name='reduction_pad_1_%s' % block_id)(h)
with backend.name_scope('block_1'):
x1_1 = _separable_conv_block(h,
filters, (5, 5),
strides=(2, 2),
block_id='reduction_left1_%s' %
block_id)
x1_2 = _separable_conv_block(p,
filters, (7, 7),
strides=(2, 2),
block_id='reduction_right1_%s' %
block_id)
x1 = layers.add([x1_1, x1_2], name='reduction_add_1_%s' % block_id)
with backend.name_scope('block_2'):
x2_1 = layers.MaxPooling2D(
(3, 3),
strides=(2, 2),
padding='valid',
name='reduction_left2_%s' % block_id)(h3)
x2_2 = _separable_conv_block(p,
filters, (7, 7),
strides=(2, 2),
block_id='reduction_right2_%s' %
block_id)
x2 = layers.add([x2_1, x2_2], name='reduction_add_2_%s' % block_id)
with backend.name_scope('block_3'):
x3_1 = layers.AveragePooling2D(
(3, 3),
strides=(2, 2),
padding='valid',
name='reduction_left3_%s' % block_id)(h3)
x3_2 = _separable_conv_block(p,
filters, (5, 5),
strides=(2, 2),
block_id='reduction_right3_%s' %
block_id)
x3 = layers.add([x3_1, x3_2], name='reduction_add3_%s' % block_id)
with backend.name_scope('block_4'):
x4 = layers.AveragePooling2D(
(3, 3),
strides=(1, 1),
padding='same',
name='reduction_left4_%s' % block_id)(x1)
x4 = layers.add([x2, x4])
with backend.name_scope('block_5'):
x5_1 = _separable_conv_block(x1,
filters, (3, 3),
block_id='reduction_left4_%s' %
block_id)
x5_2 = layers.MaxPooling2D(
(3, 3),
strides=(2, 2),
padding='valid',
name='reduction_right5_%s' % block_id)(h3)
x5 = layers.add([x5_1, x5_2], name='reduction_add4_%s' % block_id)
x = layers.concatenate([x2, x3, x4, x5],
axis=channel_dim,
name='reduction_concat_%s' % block_id)
return x, ip
def residual_inverted(
inputs: Optional[tf.Tensor] = None,
filters_in: Optional[int] = None,
filters_out: Optional[int] = None,
expand_ratio: float = 2.,
se_ratio: float = 0.25,
sigmoid_gating: bool = False,
batchnorm: bool = True,
batchnorm_kw: Optional[Dict[str, Any]] = None,
dropout: float = 0.0,
kernel_size: Tuple[int, int] = (3, 3),
order: Literal['baw', 'wba'] = 'wba',
strides: Tuple[int, int] = (1, 1),
activation: Callable[[tf.Tensor], tf.Tensor] = tf.nn.swish,
skip_mode: Literal['add', 'concat'] = 'add',
skip_ratio: float = 1.0,
name: str = 'residual_inverted',
**conv_kw,
) -> Union[tf.Tensor, ResidualSequential]:
if batchnorm_kw is None:
batchnorm_kw = {}
assert expand_ratio >= 1, (
f'Inverted residual only support expand_ratio >= 1, given {expand_ratio}'
)
## prepare
layers = []
batchnorm_kw = dict(axis=3, **batchnorm_kw)
if filters_in is None:
filters_in = inputs.shape[-1] # assume NHWC
filters = max(1, int(expand_ratio * filters_in))
if filters_out is None:
filters_out = filters_in
use_bias = not batchnorm
if np.any(np.asarray(strides) >= 2):
layers.append(
ZeroPadding2D(padding=correct_pad(inputs, kernel_size),
name=f'{name}_pad'))
pad_mode = 'valid'
else:
pad_mode = 'same'
## expand
if order == 'baw':
if batchnorm:
layers.append(
BatchNormalization(**batchnorm_kw, name=f'{name}_bn1'))
layers.append(Activation(activation, name=f'{name}_act1'))
layers.append(
Conv2D(filters=filters,
kernel_size=kernel_size,
padding=pad_mode,
strides=strides,
use_bias=use_bias,
name=f'{name}_conv1',
**conv_kw))
if order == 'wba':
if batchnorm:
layers.append(
BatchNormalization(**batchnorm_kw, name=f'{name}_bn1'))
layers.append(Activation(activation, name=f'{name}_act1'))
## squeeze
if order == 'baw':
if batchnorm:
layers.append(
BatchNormalization(**batchnorm_kw, name=f'{name}_bn2'))
layers.append(Activation(activation, name=f'{name}_act2'))
layers.append(
DepthwiseConv2D(kernel_size=kernel_size,
name=f'{name}_dwconv1',
**conv_kw))
if order == 'wba':
if batchnorm:
layers.append(
BatchNormalization(**batchnorm_kw, name=f'{name}_bn2'))
layers.append(Activation(activation, name=f'{name}_act2'))
## final
if se_ratio:
layers.append(
SqueezeExcitation(se_ratio=se_ratio,
activation=activation,
name=f'{name}_se'))
layers.append(
Conv2D(filters=filters_out * (2 if sigmoid_gating else 1),
kernel_size=(1, 1),
use_bias=use_bias,
name=f'{name}_proj1',
**conv_kw))
if batchnorm:
layers.append(BatchNormalization(**batchnorm_kw, name=f'{name}_bn3'))
if sigmoid_gating:
layers.append(SigmoidGating(name=f'{name}_gating'))
# no residual connection if strides > 1
if filters_out == filters_in and np.all(np.asarray(strides) == 1):
if dropout > 0:
layers.append(dropout2D(rate=dropout, name=f'{name}_drop'))
else:
skip_mode = 'none'
## final layer
res = ResidualSequential(skip_mode=skip_mode,
skip_ratio=skip_ratio,
layers=layers,
name=name)
if inputs is None:
return res
return res(inputs)
def block(inputs,
activation='swish',
drop_rate=0.,
name='',
filters_in=32,
filters_out=16,
kernel_size=3,
strides=1,
expand_ratio=1,
se_ratio=0.,
id_skip=True):
"""An inverted residual block.
Args:
inputs: input tensor.
activation: activation function.
drop_rate: float between 0 and 1, fraction of the input units to drop.
name: string, block label.
filters_in: integer, the number of input filters.
filters_out: integer, the number of output filters.
kernel_size: integer, the dimension of the convolution window.
strides: integer, the stride of the convolution.
expand_ratio: integer, scaling coefficient for the input filters.
se_ratio: float between 0 and 1, fraction to squeeze the input filters.
id_skip: boolean.
Returns:
output tensor for the block.
"""
bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1
# Expansion phase
filters = filters_in * expand_ratio
if expand_ratio != 1:
x = layers.Conv2D(filters,
1,
padding='same',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name + 'expand_conv')(inputs)
x = layers.BatchNormalization(axis=bn_axis, name=name + 'expand_bn')(x)
x = layers.Activation(activation, name=name + 'expand_activation')(x)
else:
x = inputs
# Depthwise Convolution
if strides == 2:
x = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(
x, kernel_size),
name=name + 'dwconv_pad')(x)
conv_pad = 'valid'
else:
conv_pad = 'same'
x = layers.DepthwiseConv2D(kernel_size,
strides=strides,
padding=conv_pad,
use_bias=False,
depthwise_initializer=CONV_KERNEL_INITIALIZER,
name=name + 'dwconv')(x)
x = layers.BatchNormalization(axis=bn_axis, name=name + 'bn')(x)
x = layers.Activation(activation, name=name + 'activation')(x)
# Squeeze and Excitation phase
if 0 < se_ratio <= 1:
filters_se = max(1, int(filters_in * se_ratio))
se = layers.GlobalAveragePooling2D(name=name + 'se_squeeze')(x)
if bn_axis == 1:
se_shape = (filters, 1, 1)
else:
se_shape = (1, 1, filters)
se = layers.Reshape(se_shape, name=name + 'se_reshape')(se)
se = layers.Conv2D(filters_se,
1,
padding='same',
activation=activation,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name + 'se_reduce')(se)
se = layers.Conv2D(filters,
1,
padding='same',
activation='sigmoid',
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name + 'se_expand')(se)
x = layers.multiply([x, se], name=name + 'se_excite')
# Output phase
x = layers.Conv2D(filters_out,
1,
padding='same',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name + 'project_conv')(x)
x = layers.BatchNormalization(axis=bn_axis, name=name + 'project_bn')(x)
if id_skip and strides == 1 and filters_in == filters_out:
if drop_rate > 0:
x = layers.Dropout(drop_rate,
noise_shape=(None, 1, 1, 1),
name=name + 'drop')(x)
x = layers.add([x, inputs], name=name + 'add')
return x
def EfficientNet(width_coefficient,
depth_coefficient,
default_size,
dropout_rate=0.2,
drop_connect_rate=0.2,
depth_divisor=8,
activation='swish',
blocks_args='default',
model_name='efficientnet',
include_top=True,
weights='imagenet',
input_tensor=None,
input_shape=None,
pooling=None,
classes=1000,
classifier_activation='softmax'):
"""Instantiates the EfficientNet architecture using given scaling coefficients.
Args:
width_coefficient: float, scaling coefficient for network width.
depth_coefficient: float, scaling coefficient for network depth.
default_size: integer, default input image size.
dropout_rate: float, dropout rate before final classifier layer.
drop_connect_rate: float, dropout rate at skip connections.
depth_divisor: integer, a unit of network width.
activation: activation function.
blocks_args: list of dicts, parameters to construct block modules.
model_name: string, model name.
include_top: whether to include the fully-connected
layer at the top of the network.
weights: one of `None` (random initialization),
'imagenet' (pre-training on ImageNet),
or the path to the weights file to be loaded.
input_tensor: optional Keras tensor
(i.e. output of `layers.Input()`)
to use as image input for the model.
input_shape: optional shape tuple, only to be specified
if `include_top` is False.
It should have exactly 3 inputs channels.
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.
classifier_activation: A `str` or callable. The activation function to use
on the "top" layer. Ignored unless `include_top=True`. Set
`classifier_activation=None` to return the logits of the "top" layer.
Returns:
A `keras.Model` instance.
Raises:
ValueError: in case of invalid argument for `weights`,
or invalid input shape.
ValueError: if `classifier_activation` is not `softmax` or `None` when
using a pretrained top layer.
"""
if blocks_args == 'default':
blocks_args = DEFAULT_BLOCKS_ARGS
if not (weights in {'imagenet', None} or file_io.file_exists_v2(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')
# Determine proper input shape
input_shape = imagenet_utils.obtain_input_shape(
input_shape,
default_size=default_size,
min_size=32,
data_format=backend.image_data_format(),
require_flatten=include_top,
weights=weights)
if input_tensor is None:
img_input = layers.Input(shape=input_shape)
else:
if not backend.is_keras_tensor(input_tensor):
img_input = layers.Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1
def round_filters(filters, divisor=depth_divisor):
"""Round number of filters based on depth multiplier."""
filters *= width_coefficient
new_filters = max(divisor,
int(filters + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_filters < 0.9 * filters:
new_filters += divisor
return int(new_filters)
def round_repeats(repeats):
"""Round number of repeats based on depth multiplier."""
return int(math.ceil(depth_coefficient * repeats))
# Build stem
x = img_input
x = layers.Rescaling(1. / 255.)(x)
x = layers.Normalization(axis=bn_axis)(x)
x = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(x, 3),
name='stem_conv_pad')(x)
x = layers.Conv2D(round_filters(32),
3,
strides=2,
padding='valid',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name='stem_conv')(x)
x = layers.BatchNormalization(axis=bn_axis, name='stem_bn')(x)
x = layers.Activation(activation, name='stem_activation')(x)
# Build blocks
blocks_args = copy.deepcopy(blocks_args)
b = 0
blocks = float(sum(round_repeats(args['repeats']) for args in blocks_args))
for (i, args) in enumerate(blocks_args):
assert args['repeats'] > 0
# Update block input and output filters based on depth multiplier.
args['filters_in'] = round_filters(args['filters_in'])
args['filters_out'] = round_filters(args['filters_out'])
for j in range(round_repeats(args.pop('repeats'))):
# The first block needs to take care of stride and filter size increase.
if j > 0:
args['strides'] = 1
args['filters_in'] = args['filters_out']
x = block(x,
activation,
drop_connect_rate * b / blocks,
name='block{}{}_'.format(i + 1, chr(j + 97)),
**args)
b += 1
# Build top
x = layers.Conv2D(round_filters(1280),
1,
padding='same',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name='top_conv')(x)
x = layers.BatchNormalization(axis=bn_axis, name='top_bn')(x)
x = layers.Activation(activation, name='top_activation')(x)
if include_top:
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
if dropout_rate > 0:
x = layers.Dropout(dropout_rate, name='top_dropout')(x)
imagenet_utils.validate_activation(classifier_activation, weights)
x = layers.Dense(classes,
activation=classifier_activation,
kernel_initializer=DENSE_KERNEL_INITIALIZER,
name='predictions')(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
elif pooling == 'max':
x = layers.GlobalMaxPooling2D(name='max_pool')(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = layer_utils.get_source_inputs(input_tensor)
else:
inputs = img_input
# Create model.
model = training.Model(inputs, x, name=model_name)
# Load weights.
if weights == 'imagenet':
if include_top:
file_suffix = '.h5'
file_hash = WEIGHTS_HASHES[model_name[-2:]][0]
else:
file_suffix = '_notop.h5'
file_hash = WEIGHTS_HASHES[model_name[-2:]][1]
file_name = model_name + file_suffix
weights_path = data_utils.get_file(file_name,
BASE_WEIGHTS_PATH + file_name,
cache_subdir='models',
file_hash=file_hash)
model.load_weights(weights_path)
elif weights is not None:
model.load_weights(weights)
return model
def MobileNetV2(input_shape=None,
alpha=1.0,
include_top=True,
weights='imagenet',
input_tensor=None,
pooling=None,
classes=1000,
classifier_activation='softmax',
**kwargs):
"""Instantiates the MobileNetV2 architecture.
Reference:
- [MobileNetV2: Inverted Residuals and Linear Bottlenecks](
https://arxiv.org/abs/1801.04381) (CVPR 2018)
Optionally loads weights pre-trained on ImageNet.
Caution: Be sure to properly pre-process your inputs to the application.
Please see `applications.mobilenet_v2.preprocess_input` for an example.
Arguments:
input_shape: Optional shape tuple, to be specified if you would
like to use a model with an input image resolution that is not
(224, 224, 3).
It should have exactly 3 inputs channels (224, 224, 3).
You can also omit this option if you would like
to infer input_shape from an input_tensor.
If you choose to include both input_tensor and input_shape then
input_shape will be used if they match, if the shapes
do not match then we will throw an error.
E.g. `(160, 160, 3)` would be one valid value.
alpha: Float between 0 and 1. controls the width of the network.
This is known as the width multiplier in the MobileNetV2 paper,
but the name is kept for consistency with `applications.MobileNetV1`
model in Keras.
- 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.
include_top: Boolean, whether to include the fully-connected
layer at the top of the network. Defaults to `True`.
weights: String, one of `None` (random initialization),
'imagenet' (pre-training on ImageNet),
or the path to the weights file to be loaded.
input_tensor: Optional Keras tensor (i.e. output of
`layers.Input()`)
to use as image input for the model.
pooling: String, 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 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.
- `max` means that global max pooling will
be applied.
classes: Integer, optional number of classes to classify images
into, only to be specified if `include_top` is True, and
if no `weights` argument is specified.
classifier_activation: A `str` or callable. The activation function to use
on the "top" layer. Ignored unless `include_top=True`. Set
`classifier_activation=None` to return the logits of the "top" layer.
**kwargs: For backwards compatibility only.
Returns:
A `keras.Model` instance.
Raises:
ValueError: in case of invalid argument for `weights`,
or invalid input shape or invalid alpha, rows when
weights='imagenet'
ValueError: if `classifier_activation` is not `softmax` or `None` when
using a pretrained top layer.
"""
global layers
if 'layers' in kwargs:
layers = kwargs.pop('layers')
else:
layers = VersionAwareLayers()
if kwargs:
raise ValueError('Unknown argument(s): %s' % (kwargs,))
if not (weights in {'imagenet', None} or file_io.file_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')
# Determine proper input shape and default size.
# If both input_shape and input_tensor are used, they should match
if input_shape is not None and input_tensor is not None:
try:
is_input_t_tensor = backend.is_keras_tensor(input_tensor)
except ValueError:
try:
is_input_t_tensor = backend.is_keras_tensor(
layer_utils.get_source_inputs(input_tensor))
except ValueError:
raise ValueError('input_tensor: ', input_tensor,
'is not type input_tensor')
if is_input_t_tensor:
if backend.image_data_format == 'channels_first':
if backend.int_shape(input_tensor)[1] != input_shape[1]:
raise ValueError('input_shape: ', input_shape, 'and input_tensor: ',
input_tensor,
'do not meet the same shape requirements')
else:
if backend.int_shape(input_tensor)[2] != input_shape[1]:
raise ValueError('input_shape: ', input_shape, 'and input_tensor: ',
input_tensor,
'do not meet the same shape requirements')
else:
raise ValueError('input_tensor specified: ', input_tensor,
'is not a keras tensor')
# If input_shape is None, infer shape from input_tensor
if input_shape is None and input_tensor is not None:
try:
backend.is_keras_tensor(input_tensor)
except ValueError:
raise ValueError('input_tensor: ', input_tensor, 'is type: ',
type(input_tensor), 'which is not a valid type')
if input_shape is None and not backend.is_keras_tensor(input_tensor):
default_size = 224
elif input_shape is None and backend.is_keras_tensor(input_tensor):
if backend.image_data_format() == 'channels_first':
rows = backend.int_shape(input_tensor)[2]
cols = backend.int_shape(input_tensor)[3]
else:
rows = backend.int_shape(input_tensor)[1]
cols = backend.int_shape(input_tensor)[2]
if rows == cols and rows in [96, 128, 160, 192, 224]:
default_size = rows
else:
default_size = 224
# If input_shape is None and no input_tensor
elif input_shape is None:
default_size = 224
# If input_shape is not None, assume default size
else:
if backend.image_data_format() == 'channels_first':
rows = input_shape[1]
cols = input_shape[2]
else:
rows = input_shape[0]
cols = input_shape[1]
if rows == cols and rows in [96, 128, 160, 192, 224]:
default_size = rows
else:
default_size = 224
input_shape = imagenet_utils.obtain_input_shape(
input_shape,
default_size=default_size,
min_size=32,
data_format=backend.image_data_format(),
require_flatten=include_top,
weights=weights)
if backend.image_data_format() == 'channels_last':
row_axis, col_axis = (0, 1)
else:
row_axis, col_axis = (1, 2)
rows = input_shape[row_axis]
cols = input_shape[col_axis]
if weights == 'imagenet':
if alpha not in [0.35, 0.50, 0.75, 1.0, 1.3, 1.4]:
raise ValueError('If imagenet weights are being loaded, '
'alpha can be one of `0.35`, `0.50`, `0.75`, '
'`1.0`, `1.3` or `1.4` only.')
if rows != cols or rows not in [96, 128, 160, 192, 224]:
rows = 224
logging.warning('`input_shape` is undefined or non-square, '
'or `rows` is not in [96, 128, 160, 192, 224].'
' Weights for input shape (224, 224) will be'
' loaded as the default.')
if input_tensor is None:
img_input = layers.Input(shape=input_shape)
else:
if not backend.is_keras_tensor(input_tensor):
img_input = layers.Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1
first_block_filters = _make_divisible(32 * alpha, 8)
x = layers.ZeroPadding2D(
padding=imagenet_utils.correct_pad(img_input, 3),
name='Conv1_pad')(img_input)
x = layers.Conv2D(
first_block_filters,
kernel_size=3,
strides=(2, 2),
padding='valid',
use_bias=False,
name='Conv1')(
x)
x = layers.BatchNormalization(
axis=channel_axis, epsilon=1e-3, momentum=0.999, name='bn_Conv1')(
x)
x = layers.ReLU(6., name='Conv1_relu')(x)
x = _inverted_res_block(
x, filters=16, alpha=alpha, stride=1, expansion=1, block_id=0)
x = _inverted_res_block(
x, filters=24, alpha=alpha, stride=2, expansion=6, block_id=1)
x = _inverted_res_block(
x, filters=24, alpha=alpha, stride=1, expansion=6, block_id=2)
x = _inverted_res_block(
x, filters=32, alpha=alpha, stride=2, expansion=6, block_id=3)
x = _inverted_res_block(
x, filters=32, alpha=alpha, stride=1, expansion=6, block_id=4)
x = _inverted_res_block(
x, filters=32, alpha=alpha, stride=1, expansion=6, block_id=5)
x = _inverted_res_block(
x, filters=64, alpha=alpha, stride=2, expansion=6, block_id=6)
x = _inverted_res_block(
x, filters=64, alpha=alpha, stride=1, expansion=6, block_id=7)
x = _inverted_res_block(
x, filters=64, alpha=alpha, stride=1, expansion=6, block_id=8)
x = _inverted_res_block(
x, filters=64, alpha=alpha, stride=1, expansion=6, block_id=9)
x = _inverted_res_block(
x, filters=96, alpha=alpha, stride=1, expansion=6, block_id=10)
x = _inverted_res_block(
x, filters=96, alpha=alpha, stride=1, expansion=6, block_id=11)
x = _inverted_res_block(
x, filters=96, alpha=alpha, stride=1, expansion=6, block_id=12)
x = _inverted_res_block(
x, filters=160, alpha=alpha, stride=2, expansion=6, block_id=13)
x = _inverted_res_block(
x, filters=160, alpha=alpha, stride=1, expansion=6, block_id=14)
x = _inverted_res_block(
x, filters=160, alpha=alpha, stride=1, expansion=6, block_id=15)
x = _inverted_res_block(
x, 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 = layers.Conv2D(
last_block_filters, kernel_size=1, use_bias=False, name='Conv_1')(
x)
x = layers.BatchNormalization(
axis=channel_axis, epsilon=1e-3, momentum=0.999, name='Conv_1_bn')(
x)
x = layers.ReLU(6., name='out_relu')(x)
if include_top:
x = layers.GlobalAveragePooling2D()(x)
imagenet_utils.validate_activation(classifier_activation, weights)
x = layers.Dense(classes, activation=classifier_activation,
name='predictions')(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D()(x)
elif pooling == 'max':
x = layers.GlobalMaxPooling2D()(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = layer_utils.get_source_inputs(input_tensor)
else:
inputs = img_input
# Create model.
model = training.Model(inputs, x, name='mobilenetv2_%0.2f_%s' % (alpha, rows))
# Load weights.
if weights == 'imagenet':
if include_top:
model_name = ('mobilenet_v2_weights_tf_dim_ordering_tf_kernels_' +
str(alpha) + '_' + str(rows) + '.h5')
weight_path = BASE_WEIGHT_PATH + model_name
weights_path = data_utils.get_file(
model_name, weight_path, cache_subdir='models')
else:
model_name = ('mobilenet_v2_weights_tf_dim_ordering_tf_kernels_' +
str(alpha) + '_' + str(rows) + '_no_top' + '.h5')
weight_path = BASE_WEIGHT_PATH + model_name
weights_path = data_utils.get_file(
model_name, weight_path, cache_subdir='models')
model.load_weights(weights_path)
elif weights is not None:
model.load_weights(weights)
return model
def build_model(self):
# Build stem
x = self.img_input
x = layers.Rescaling(1. / 255.)(x)
x = layers.Normalization(axis=self.bn_axis)(x)
x = tf.keras.layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(
x, 3),
name='stem_conv_pad')(x)
x = tf.keras.layers.Conv2D(self.round_filters(32),
3,
strides=2,
padding='valid',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name='stem_conv')(x)
x = build_normalization(**self.normalization, name='stem_bn')(x)
x = build_activation(**self.activation, name='stem_activation')(x)
block_outputs = [x]
# Build blocks
blocks_args = copy.deepcopy(self.blocks_args)
b = 0
blocks = float(sum(args['repeats'] for args in blocks_args))
output_stages = [1, 2, 4, 6]
for (i, args) in enumerate(blocks_args):
assert args['repeats'] > 0
# Update block input and output filters based on depth multiplier.
args['filters_in'] = self.round_filters(args['filters_in'])
args['filters_out'] = self.round_filters(args['filters_out'])
strides = args["strides"]
for j in range(self.round_repeats(args.pop('repeats'))):
# The first block needs to take care of stride and filter size increase.
if j > 0:
args['strides'] = 1
args['filters_in'] = args['filters_out']
x = self.block(x,
self.activation,
self.drop_connect_rate * b / blocks,
name='block{}{}_'.format(i + 1, chr(j + 97)),
**args)
b += 1
if i in output_stages:
block_outputs.append(x)
# Build top
x = tf.keras.layers.Conv2D(self.round_filters(1280),
1,
padding='same',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name='top_conv')(x)
x = build_normalization(**self.normalization, name='top_bn')(x)
x = build_activation(**self.activation, name='top_activation')(x)
if -1 in self.output_indices:
x = tf.keras.layers.GlobalAveragePooling2D(name='avg_pool')(x)
if self.dropout_rate > 0:
x = tf.keras.layers.Dropout(self.dropout_rate,
name='top_dropout')(x)
x = tf.keras.layers.Dense(
1000,
activation=self.classifier_activation,
kernel_initializer=DENSE_KERNEL_INITIALIZER,
name='predictions')(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if self.input_tensor is not None:
inputs = tf.keras.utils.get_source_inputs(self.input_tensor)
else:
inputs = self.img_input
# Create model.
return tf.keras.Model(inputs=self.img_input,
outputs=x,
name=model_name)
return [block_outputs[i - 1] for i in self.output_indices]
def EfficientNet(width_coefficient,
depth_coefficient,
default_size,
dropout_rate=0.2,
drop_connect_rate=0.2,
depth_divisor=8,
activation='swish',
blocks_args='default',
model_name='efficientnet',
include_top=True,
weights='imagenet',
input_tensor=None,
input_shape=None,
pooling=None,
classes=1000,
classifier_activation='softmax'):
if blocks_args == 'default':
blocks_args = DEFAULT_BLOCKS_ARGS
if input_tensor is None:
img_input = layers.Input(shape=input_shape)
else:
if not backend.is_keras_tensor(input_tensor):
img_input = layers.Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1
def round_filters(filters, divisor=depth_divisor):
"""Round number of filters based on depth multiplier."""
filters *= width_coefficient
new_filters = max(divisor,
int(filters + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_filters < 0.9 * filters:
new_filters += divisor
return int(new_filters)
def round_repeats(repeats):
"""Round number of repeats based on depth multiplier."""
return int(math.ceil(depth_coefficient * repeats))
# Build stem
x = img_input
x = layers.Rescaling(1. / 255.)(x)
x = layers.Normalization(axis=bn_axis)(x)
x = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(x, 3),
name='stem_conv_pad')(x)
x = layers.Conv2D(round_filters(32),
3,
strides=2,
padding='valid',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name='stem_conv')(x)
x = layers.BatchNormalization(axis=bn_axis, name='stem_bn')(x)
x = layers.Activation(activation, name='stem_activation')(x)
# Build blocks
blocks_args = copy.deepcopy(blocks_args)
b = 0
blocks = float(sum(round_repeats(args['repeats']) for args in blocks_args))
for (i, args) in enumerate(blocks_args):
assert args['repeats'] > 0
# Update block input and output filters based on depth multiplier.
args['filters_in'] = round_filters(args['filters_in'])
args['filters_out'] = round_filters(args['filters_out'])
for j in range(round_repeats(args.pop('repeats'))):
# The first block needs to take care of stride and filter size increase.
if j > 0:
args['strides'] = 1
args['filters_in'] = args['filters_out']
x = block(x,
activation,
drop_connect_rate * b / blocks,
name='block{}{}_'.format(i + 1, chr(j + 97)),
**args)
b += 1
# Build top
x = layers.Conv2D(round_filters(1280),
1,
padding='same',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name='top_conv')(x)
x = layers.BatchNormalization(axis=bn_axis, name='top_bn')(x)
x = layers.Activation(activation, name='top_activation')(x)
if include_top:
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
if dropout_rate > 0:
x = layers.Dropout(dropout_rate, name='top_dropout')(x)
imagenet_utils.validate_activation(classifier_activation, weights)
x = layers.Dense(classes,
activation=classifier_activation,
kernel_initializer=DENSE_KERNEL_INITIALIZER,
name='predictions')(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
elif pooling == 'max':
x = layers.GlobalMaxPooling2D(name='max_pool')(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = layer_utils.get_source_inputs(input_tensor)
else:
inputs = img_input
# Create model.
model = training.Model(inputs, x, name=model_name)
return model
def MobileNetV2(input_shape=None, alpha=1.0):
if input_shape is None:
default_size = 224
else:
if backend.image_data_format() == 'channels_first':
rows = input_shape[1]
cols = input_shape[2]
else:
rows = input_shape[0]
cols = input_shape[1]
if rows == cols and rows in [96, 128, 160, 192, 224]:
default_size = rows
else:
default_size = 224
input_shape = imagenet_utils.obtain_input_shape(
input_shape,
default_size=default_size,
min_size=32,
data_format=backend.image_data_format(),
require_flatten=True)
row_axis = 0 if backend.image_data_format() == 'channels_last' else 1
rows = input_shape[row_axis]
img_input = layers.Input(shape=input_shape)
channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1
first_block_filters = _make_divisible(32 * alpha, 8)
x = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(img_input, 3),
name='Conv1_pad')(img_input)
x = layers.Conv2D(first_block_filters,
kernel_size=3,
strides=(2, 2),
padding='valid',
use_bias=False,
name='Conv1')(x)
x = layers.BatchNormalization(axis=channel_axis,
epsilon=1e-3,
momentum=0.999,
name='bn_Conv1')(x)
x = layers.ReLU(6., name='Conv1_relu')(x)
x = _inverted_res_block(x,
filters=16,
alpha=alpha,
stride=1,
expansion=1,
block_id=0)
x = _inverted_res_block(x,
filters=24,
alpha=alpha,
stride=2,
expansion=6,
block_id=1)
x = _inverted_res_block(x,
filters=24,
alpha=alpha,
stride=1,
expansion=6,
block_id=2)
x = _inverted_res_block(x,
filters=32,
alpha=alpha,
stride=2,
expansion=6,
block_id=3)
x = _inverted_res_block(x,
filters=32,
alpha=alpha,
stride=1,
expansion=6,
block_id=4)
x = _inverted_res_block(x,
filters=32,
alpha=alpha,
stride=1,
expansion=6,
block_id=5)
x = _inverted_res_block(x,
filters=64,
alpha=alpha,
stride=2,
expansion=6,
block_id=6)
x = _inverted_res_block(x,
filters=64,
alpha=alpha,
stride=1,
expansion=6,
block_id=7)
x = _inverted_res_block(x,
filters=64,
alpha=alpha,
stride=1,
expansion=6,
block_id=8)
x = _inverted_res_block(x,
filters=64,
alpha=alpha,
stride=1,
expansion=6,
block_id=9)
x = _inverted_res_block(x,
filters=96,
alpha=alpha,
stride=1,
expansion=6,
block_id=10)
x = _inverted_res_block(x,
filters=96,
alpha=alpha,
stride=1,
expansion=6,
block_id=11)
x = _inverted_res_block(x,
filters=96,
alpha=alpha,
stride=1,
expansion=6,
block_id=12)
x = _inverted_res_block(x,
filters=160,
alpha=alpha,
stride=2,
expansion=6,
block_id=13)
x = _inverted_res_block(x,
filters=160,
alpha=alpha,
stride=1,
expansion=6,
block_id=14)
x = _inverted_res_block(x,
filters=160,
alpha=alpha,
stride=1,
expansion=6,
block_id=15)
x = _inverted_res_block(x,
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 = layers.Conv2D(last_block_filters,
kernel_size=1,
use_bias=False,
name='Conv_1')(x)
x = layers.BatchNormalization(axis=channel_axis,
epsilon=1e-3,
momentum=0.999,
name='Conv_1_bn')(x)
x = layers.ReLU(6., name='out_relu')(x)
x = layers.GlobalAveragePooling2D()(x)
imagenet_utils.validate_activation('softmax', None)
x = layers.Dense(NUM_CLASSES, activation='softmax', name='predictions')(x)
# Create model.
model = training.Model(img_input,
x,
name='mobilenetv2_%0.2f_%s' % (alpha, rows))
return model
def mbconvblock(inputs,
activations='swish',
droprate=0.,
name='',
filters_in=32,
filters_out=16,
kernel_size=3,
strides=1,
expand_ratio=1,
se_ratio=0.,
id_skip=True,
attn=True):
bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1
#expansion phase
filters = filters_in * expand_ratio
if expand_ratio != 1:
x = layers.Conv2D(filters,
kernel_size=1,
padding='same',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER)(inputs)
x = layers.BatchNormalization(axis=bn_axis)(x)
x = layers.Activation(activations)(x) #swish activation
else:
x = inputs
#depthwise conv
if strides == 2:
x = layers.ZeroPadding2D(
padding=imagenet_utils.correct_pad(x, kernel_size))(x)
conv_pad = 'valid'
else:
conv_pad = 'same'
x = layers.DepthwiseConv2D(
kernel_size=kernel_size,
strides=strides,
padding=conv_pad,
use_bias=False,
depthwise_initializer=CONV_KERNEL_INITIALIZER)(x)
x = layers.BatchNormalization(axis=bn_axis)(x)
x = layers.Activation(activations)(x)
#squeeze and excitation
if 0 < se_ratio <= 1:
filters_se = max(1, int(filters_in * se_ratio))
se = layers.GlobalAveragePooling2D()(x)
se = layers.Reshape((1, 1, filters))(se)
se = layers.Conv2D(filters_se,
kernel_size=1,
padding='same',
activation=activations,
kernel_initializer=CONV_KERNEL_INITIALIZER)(
se) #reduce
se = layers.Conv2D(filters,
kernel_size=1,
padding='same',
activation='sigmoid',
kernel_initializer=CONV_KERNEL_INITIALIZER)(
se) #expand
x = layers.multiply([x, se])
if attn:
avgp = layers.Lambda(lambda x: K.mean(x, axis=3, keepdims=True))(x)
maxp = layers.Lambda(lambda x: K.max(x, axis=3, keepdims=True))(x)
conc = layers.Concatenate(axis=3)([avgp, maxp])
attn_mask = layers.Conv2D(
filters=1,
kernel_size=1,
activation='sigmoid',
padding='same',
name=name + '_Attn',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER)(conc)
#Output phase
x = layers.Conv2D(filters_out,
kernel_size=1,
padding='same',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER)(x)
x = layers.BatchNormalization(axis=bn_axis)(x)
if id_skip and strides == 1 and filters_in == filters_out:
if droprate > 0:
x = layers.Dropout(droprate, noise_shape=(None, 1, 1, 1))(x)
x = layers.add([x, inputs])
if attn:
x = layers.multiply([x, attn_mask])
return x
def efficientnetB0(width_coeff=1,
depth_coeff=1,
default_img_size=224,
dropoutrate=0.2,
drop_connect_rate=0.2,
depth_divisor=8,
activations='swish',
block_args='default',
include_top=True,
weights='imagenet',
input_tensor=None,
input_shape=None,
pooling=None,
classes=1000,
classifier_activation='softmax'):
if block_args == 'default':
block_args = DEFAULT_BLOCKS_ARGS
if not (weights in {'imagenet', None} or file_io.file_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')
# proper input shape check
input_shape = imagenet_utils.obtain_input_shape(
input_shape,
default_size=default_img_size,
min_size=32,
data_format=backend.image_data_format(),
require_flatten=include_top,
weights=weights)
if input_tensor is None:
img_input = layers.Input(shape=input_shape)
else:
if not backend.is_keras_tensor(input_tensor):
img_input = layers.Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1
def round_filters(filters, divisor=depth_divisor):
""" round number of filters based on depth multiplier"""
filters *= width_coeff
new_filters = max(divisor,
int(filters + divisor / 2) // divisor * divisor)
#making sure round down does not go down by 10%
if new_filters < 0.9 * filters:
new_filters += divisor
return int(new_filters)
def round_repeats(repeats):
"""Round number of repeats based on depth multiplier."""
return int(math.ceil(depth_coeff * repeats))
#building stem
x = img_input
x = layers.Rescaling(1. / 255.)(x)
x = layers.Normalization(axis=bn_axis)(x)
x = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(x, 3))(x)
x = layers.Conv2D(round_filters(32),
kernel_size=3,
strides=2,
padding='valid',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER)(x)
x = layers.BatchNormalization(axis=bn_axis)(x)
x = layers.Activation(activations)(x)
#building blocks
block_args = copy.deepcopy(block_args)
b = 0
blocks = float(sum(round_repeats(args['repeats']) for args in block_args))
for (i, args) in enumerate(block_args):
assert args['repeats'] > 0
args['filters_in'] = round_filters(args['filters_in'])
args['filters_out'] = round_filters(args['filters_out'])
for j in range(round_repeats(args.pop('repeats'))):
# The first block needs to take care of stride and filter size increase.
if j > 0:
args['strides'] = 1
args['filters_in'] = args['filters_out']
x = mbconvblock(inputs=x,
activations=activations,
droprate=drop_connect_rate * b / blocks,
name=str(i),
**args)
b += 1
#build top
x = layers.Conv2D(round_filters(1280),
kernel_size=1,
padding='same',
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER)(x)
x = layers.BatchNormalization(axis=bn_axis)(x)
x = layers.Activation(activations)(x)
if include_top:
x = layers.GlobalAveragePooling2D()(x)
if dropoutrate > 0:
x = layers.Dropout(dropoutrate)(x)
imagenet_utils.validate_activation(classifier_activation, weights)
x = layers.Dense(classes,
activation=classifier_activation,
kernel_initializer=DENSE_KERNEL_INITIALIZER)(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D()(x)
elif pooling == 'max':
x = layers.GlobalMaxPooling2D()(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = layer_utils.get_source_inputs(input_tensor)
else:
inputs = img_input
#creating model
model = training.Model(inputs, x)
return model
def mbconv_block(inputs,
global_params,
block_args,
normalization,
drop_connect_rate=None,
trainable=True,
l2_regularizer=tf.keras.regularizers.l2(l=4e-5),
name=""):
expand_ratio = block_args.expand_ratio
data_format = global_params.data_format
_momentum = global_params.batch_norm_momentum
_epsilon = global_params.batch_norm_epsilon
_axis = normalization["axis"]
_mean_axis = [1, 2] if _axis == -1 or _axis == 3 else [2, 3]
filters = block_args.in_filters * expand_ratio
expand_ratio = expand_ratio
if expand_ratio != 1:
x = tf.keras.layers.Conv2D(
filters=filters,
kernel_size=(1, 1),
strides=(1, 1),
padding="same",
data_format=data_format,
use_bias=False,
name=name + "/conv2d",
trainable=trainable,
kernel_initializer=conv2d_kernel_initializer)(inputs)
x = build_normalization(**normalization,
name=name + "/batch_normalization")(x)
x = tf.keras.layers.Activation("swish", name=name + "/swish")(x)
else:
x = inputs
# Depthwise Convolution
if block_args.strides == 2:
x = tf.keras.layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(
x, block_args.kernel_size),
name=name + '/dwconv_pad')(x)
conv_pad = 'valid'
else:
conv_pad = 'same'
x = tf.keras.layers.DepthwiseConv2D(
kernel_size=block_args.kernel_size,
strides=block_args.strides,
padding=conv_pad,
data_format=data_format,
use_bias=False,
trainable=trainable,
depthwise_initializer=conv2d_kernel_initializer,
name=name + "/depthwise_conv2d")(x)
x = build_normalization(**normalization,
name=name + "/batch_normalization" if expand_ratio
== 1 else name + "/batch_normalization_1")(x)
x = tf.keras.layers.Activation(
"swish",
name=name + "/swish" if expand_ratio == 1 else name + "/swish_1")(x)
has_se = block_args.se_ratio is not None and 0 < block_args.se_ratio < 1
if has_se:
squeeze_filters = max(1,
int(block_args.in_filters * block_args.se_ratio))
se = tf.keras.layers.Lambda(
lambda inp: tf.reduce_mean(inp, axis=_mean_axis, keepdims=True),
name=name + "/se/global_pooling")(x)
se = tf.keras.layers.Conv2D(
filters=squeeze_filters,
kernel_size=(1, 1),
strides=(1, 1),
padding="same",
data_format=data_format,
use_bias=True,
kernel_initializer=conv2d_kernel_initializer,
trainable=trainable,
name=name + "/se/conv2d")(se)
se = tf.keras.layers.Activation("swish", name=name + "/se/swish_1")(se)
se = tf.keras.layers.Conv2D(
filters=filters,
kernel_size=(1, 1),
strides=(1, 1),
padding="same",
data_format=data_format,
use_bias=True,
trainable=trainable,
kernel_initializer=conv2d_kernel_initializer,
name=name + "/se/conv2d_1")(se)
se = tf.keras.layers.Activation("sigmoid",
name=name + "/se/sigmoid")(se)
x = tf.keras.layers.Multiply(name=name + "/se/multiply")([se, x])
x = tf.keras.layers.Conv2D(
block_args.out_filters,
kernel_size=(1, 1),
strides=(1, 1),
padding="same",
data_format=data_format,
use_bias=False,
trainable=trainable,
kernel_initializer=conv2d_kernel_initializer,
name=name + "/conv2d" if expand_ratio == 1 else name + "/conv2d_1")(x)
x = build_normalization(**normalization,
name=name + "/batch_normalization_2"
if expand_ratio > 1 else name +
"/batch_normalization_1")(x)
if block_args.id_skip:
if all(s == 1 for s in block_args.strides
) and block_args.in_filters == block_args.out_filters:
# x = DropConnect(drop_connect_rate, name=name + "/drop_connect")(x)
# x = tf.keras.layers.Dropout(drop_connect_rate, noise_shape=(None, 1, 1, 1), name=name + '/drop')(x)
x = tf.keras.layers.Add(name=name + "/add")([x, inputs])
return x
def _reduction_a_cell(ip, p, filters, block_id=None):
channel_dim = -1
with backend.name_scope('reduction_A_block_%s' % block_id):
p = _adjust_block(p, ip, filters, block_id)
h = layers.Activation('relu')(ip)
h = layers.Conv2D(filters, (1, 1),
strides=(1, 1),
padding='same',
name='reduction_conv_1_%s' % block_id,
use_bias=False,
kernel_initializer='he_normal')(h)
h = layers.BatchNormalization(axis=channel_dim,
momentum=0.9997,
epsilon=1e-3,
name='reduction_bn_1_%s' % block_id)(h)
h3 = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(h, 3),
name='reduction_pad_1_%s' % block_id)(h)
with backend.name_scope('block_1'):
x1_1 = _separable_conv_block(h,
filters, (5, 5),
strides=(2, 2),
block_id='reduction_left1_%s' %
block_id)
x1_2 = _separable_conv_block(p,
filters, (7, 7),
strides=(2, 2),
block_id='reduction_right1_%s' %
block_id)
x1 = layers.add([x1_1, x1_2], name='reduction_add_1_%s' % block_id)
with backend.name_scope('block_2'):
x2_1 = layers.MaxPooling2D(
(3, 3),
strides=(2, 2),
padding='valid',
name='reduction_left2_%s' % block_id)(h3)
x2_2 = _separable_conv_block(p,
filters, (7, 7),
strides=(2, 2),
block_id='reduction_right2_%s' %
block_id)
x2 = layers.add([x2_1, x2_2], name='reduction_add_2_%s' % block_id)
with backend.name_scope('block_3'):
x3_1 = layers.AveragePooling2D(
(3, 3),
strides=(2, 2),
padding='valid',
name='reduction_left3_%s' % block_id)(h3)
x3_2 = _separable_conv_block(p,
filters, (5, 5),
strides=(2, 2),
block_id='reduction_right3_%s' %
block_id)
x3 = layers.add([x3_1, x3_2], name='reduction_add3_%s' % block_id)
with backend.name_scope('block_4'):
x4 = layers.AveragePooling2D(
(3, 3),
strides=(1, 1),
padding='same',
name='reduction_left4_%s' % block_id)(x1)
x4 = layers.add([x2, x4])
with backend.name_scope('block_5'):
x5_1 = _separable_conv_block(x1,
filters, (3, 3),
block_id='reduction_left4_%s' %
block_id)
x5_2 = layers.MaxPooling2D(
(3, 3),
strides=(2, 2),
padding='valid',
name='reduction_right5_%s' % block_id)(h3)
x5 = layers.add([x5_1, x5_2], name='reduction_add4_%s' % block_id)
x = layers.concatenate([x2, x3, x4, x5],
axis=channel_dim,
name='reduction_concat_%s' % block_id)
return x, ip
