def inverted_residual_block(x,
expanded_channels,
output_channels,
strides=1):
m = layers.Conv2D(expanded_channels, 1, padding="same",
use_bias=False)(x)
m = layers.BatchNormalization()(m)
m = tf.nn.swish(m)
if strides == 2:
m = layers.ZeroPadding2D(
padding=imagenet_utils.correct_pad(m, 3))(m)
m = layers.DepthwiseConv2D(3,
strides=strides,
padding="same" if strides == 1 else "valid",
use_bias=False)(m)
m = layers.BatchNormalization()(m)
m = tf.nn.swish(m)
m = layers.Conv2D(output_channels, 1, padding="same",
use_bias=False)(m)
m = layers.BatchNormalization()(m)
if tf.math.equal(x.shape[-1], output_channels) and strides == 1:
return layers.Add()([m, x])
return m
def _separable_conv_block(ip,
filters,
kernel_size=(3, 3),
strides=(1, 1),
block_id=None):
"""Adds 2 blocks of [relu-separable conv-batchnorm].
Args:
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)
# Ensure the number of filters on the last 1x1 convolution is divisible by 8.
pointwise_filters = _make_divisible(pointwise_conv_filters, 8)
x = inputs
prefix = 'block_{}_'.format(block_id)
if block_id:
# Expand with a pointwise 1x1 convolution.
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 3x3 convolution.
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 wiht a pointwise 1x1 convolution.
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 _reduction_a_cell(ip, p, filters, block_id=None):
"""Adds a Reduction cell for NASNet-A (Fig. 4 in the paper).
Args:
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 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 tf.io.gfile.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
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
