def __init__(self, layers=None, slic = 1, concat_axis=-1, init='glorot_uniform',
arguments=None, node_indices=None, tensor_indices=None,
name=None):
self.layers = layers
self.slic = slic
self.batch_size = 32
self.node_indices = node_indices
self.init = initializations.get(init)
self.arguments = arguments if arguments else {}
self.mode = "concat"
# Layer parameters.
self.inbound_nodes = []
self.outbound_nodes = []
self.constraints = {}
self._trainable_weights = []
self._non_trainable_weights = []
self.regularizers = []
self.supports_masking = False
self.uses_learning_phase = False
self.input_spec = None # Compatible with anything.
if not name:
prefix = self.__class__.__name__.lower()
name = prefix + '_' + str(K.get_uid(prefix))
self.name = name
if layers:
if not node_indices:
node_indices = [0 for _ in range(len(layers))]
input_shape = self._arguments_validation(layers,
node_indices, tensor_indices)
self.build(input_shape)
self.add_inbound_node(layers, node_indices, tensor_indices)
else:
self.built = False
def __init__(self, hid, layers=None, name=None):
super(Autoencoder, self).__init__()
self.layers = [] # Stack of layers.
self.model = None # Internal Model instance.
self.inputs = [] # List of input tensors
self.outputs = [] # List of length 1: the output tensor (unique).
self._trainable = True
self._initial_weights = None
self.hid = hid
# Model attributes.
self.inbound_nodes = []
self.outbound_nodes = []
self.built = False
self.pretrain = False
# Set model name.
if not name:
prefix = 'sequential_'
name = prefix + str(K.get_uid(prefix))
self.name = name
# Add to the model any layers passed to the constructor.
if layers:
for layer in layers:
self.add(layer)
def ConvNeXtBlock(
projection_dim, drop_path_rate=0.0, layer_scale_init_value=1e-6, name=None
):
"""ConvNeXt block.
References:
- https://arxiv.org/abs/2201.03545
- https://github.com/facebookresearch/ConvNeXt/blob/main/models/convnext.py
Notes:
In the original ConvNeXt implementation (linked above), the authors use
`Dense` layers for pointwise convolutions for increased efficiency.
Following that, this implementation also uses the same.
Args:
projection_dim (int): Number of filters for convolution layers. In the
ConvNeXt paper, this is referred to as projection dimension.
drop_path_rate (float): Probability of dropping paths. Should be within
[0, 1].
layer_scale_init_value (float): Layer scale value. Should be a small float
number.
name: name to path to the keras layer.
Returns:
A function representing a ConvNeXtBlock block.
"""
if name is None:
name = "prestem" + str(backend.get_uid("prestem"))
def apply(inputs):
x = inputs
x = layers.Conv2D(
filters=projection_dim,
kernel_size=7,
padding="same",
groups=projection_dim,
name=name + "_depthwise_conv",
)(x)
x = layers.LayerNormalization(epsilon=1e-6, name=name + "_layernorm")(x)
x = layers.Dense(4 * projection_dim, name=name + "_pointwise_conv_1")(x)
x = layers.Activation("gelu", name=name + "_gelu")(x)
x = layers.Dense(projection_dim, name=name + "_pointwise_conv_2")(x)
if layer_scale_init_value is not None:
x = LayerScale(
layer_scale_init_value,
projection_dim,
name=name + "_layer_scale",
)(x)
if drop_path_rate:
layer = StochasticDepth(
drop_path_rate, name=name + "_stochastic_depth"
)
else:
layer = layers.Activation("linear", name=name + "_identity")
return inputs + layer(x)
return apply
def __init__(self, inputs, name=None):
if K.backend() == 'theano' or K.image_dim_ordering() == 'th':
raise RuntimeError(
"Only support tensorflow backend or image ordering")
self.inputs = inputs
# Layer parameters
self.inbound_nodes = []
self.outbound_nodes = []
self.constraints = {}
self._trainable_weights = []
self._non_trainable_weights = []
self.supports_masking = True
self.uses_learning_phase = False
self.input_spec = None
self.trainable = False
if not name:
prefix = self.__class__.__name__.lower()
name = prefix + '_' + str(K.get_uid(prefix))
self.name = name
if inputs:
# The inputs is a bunch of nodes shares the same input.
self.built = True
# self.add_inbound_node(inputs, node_indices, tensor_indices)
else:
self.built = False
def PreStem(name=None):
"""Normalizes inputs with ImageNet-1k mean and std.
Args:
name (str): Name prefix.
Returns:
A presemt function.
"""
if name is None:
name = "prestem" + str(backend.get_uid("prestem"))
def apply(x):
x = layers.Normalization(
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
variance=[
(0.229 * 255) ** 2,
(0.224 * 255) ** 2,
(0.225 * 255) ** 2,
],
name=name + "_prestem_normalization",
)(x)
return x
return apply
def __init__(self,
size,
initializer='glorot_uniform',
regularizer=None,
name=None,
**kwargs):
self.size = tuple(size)
self.initializer = initializers.get(initializer)
self.regularizer = regularizers.get(regularizer)
if not name:
prefix = 'shared_weight'
name = prefix + '_' + str(K.get_uid(prefix))
Layer.__init__(self, name=name, **kwargs)
with K.name_scope(self.name):
self.kernel = self.add_weight(shape=self.size,
initializer=self.initializer,
name='kernel',
regularizer=self.regularizer)
self.trainable = True
self.built = True
# self.sparse = sparse
# input_tensor = self.kernel * 1.0
self.is_placeholder = False
def _shared_name(self):
"""Returns a shared name to be used by the table."""
shared_name = "NULL_INITIALIZER_"
if tf.executing_eagerly():
# Ensure a unique name when eager execution is enabled to avoid spurious
# sharing issues..
shared_name += str(backend.get_uid(shared_name))
return shared_name
def __init__(self, **kwargs):
''' Constructor of a composed layer, which should be called as the last function call of the constructor of its sub class by that sub class to construct its children layers.
Note that most of following lines are shamelessly adapted from keras
'''
# Logic copied from layer with small adaption
if not hasattr(self, 'input_spec'):
self.input_spec = None
if not hasattr(self, 'supports_masking'):
self.supports_masking = False
self._uses_learning_phase = False
# these lists will be filled via successive calls
# to self.add_inbound_node()
self.inbound_nodes = []
self.outbound_nodes = []
self._trainable_weights = []
self._non_trainable_weights = []
self._regularizers = []
self._constraints = {} # dict {tensor: constraint instance}
self.built = False
# these properties should be set by the user via keyword arguments.
# note that 'input_dtype', 'input_shape' and 'batch_input_shape'
# are only applicable to input layers: do not pass these keywords
# to non-input layers.
allowed_kwargs = {
'input_shape', 'batch_input_shape', 'input_dtype', 'name',
'trainable', 'create_input_layer'
}
for kwarg in kwargs.keys():
assert kwarg in allowed_kwargs, 'Keyword argument not understood: ' + kwarg
name = kwargs.get('name')
if not name:
prefix = self.__class__.__name__.lower()
name = prefix + '_' + str(K.get_uid(prefix))
self.name = name
self.trainable = kwargs.get('trainable', True)
if 'batch_input_shape' in kwargs or 'input_shape' in kwargs:
# in this case we will create an input layer
# to insert before the current layer
if 'batch_input_shape' in kwargs:
batch_input_shape = tuple(kwargs['batch_input_shape'])
elif 'input_shape' in kwargs:
batch_input_shape = (None, ) + tuple(kwargs['input_shape'])
self.batch_input_shape = batch_input_shape
input_dtype = kwargs.get('input_dtype', K.floatx())
self.input_dtype = input_dtype
if 'create_input_layer' in kwargs:
self.create_input_layer(batch_input_shape, input_dtype)
self._updates = []
self._stateful = False
self._layers = []
def __init__(self,
window_size,
axes,
layers=None,
name=None,
node_indices=None,
tensor_indices=None,
output_shape=None,
output_mask=None,
**kwargs):
self.window_size = window_size
self.shape_axes = axes
if axes == 1:
self.transform_axis = 2
elif axes == 2:
self.transform_axis = 1
else:
#TODO maybe throw error
self.transform_axis = axes
self._output_shape = output_shape
self._output_mask = output_mask
self.inbound_nodes = []
self.outbound_nodes = []
self.constraints = {}
self.regularizers = []
self.trainable_weights = []
self.non_trainable_weights = []
self.supports_masking = False
self.uses_learning_phase = False
self.input_spec = None # compatible with whatever
self.input_dim = None
self.input_layers = layers
self.node_indices = node_indices
self.tensor_indices = tensor_indices
self.trainable = False
if not name:
prefix = self.__class__.__name__.lower()
name = prefix + '_' + str(K.get_uid(prefix))
self.name = name
if layers:
# this exists for backwards compatibility.
# equivalent to:
# merge = Merge(layers=None)
# output = merge([input_tensor_1, input_tensor_2])
if not node_indices:
# by default we connect to
# the 1st output stream in the input layer
node_indices = [0 for _ in range(len(layers))]
self.built = True
self.add_inbound_node(self.input_layers, self.node_indices,
self.tensor_indices)
self.trainable_weights = []
def __init__(self,
input_shape=None,
batch_size=None,
batch_input_shape=None,
dtype=None,
input_tensor=None,
sparse=False,
name=None):
#self.input_spec = None
self.supports_masking = False
self.uses_learning_phase = False
self.trainable = False
self.built = True
self._inbound_nodes = []
self._outbound_nodes = []
self.trainable_weights = []
self.non_trainable_weights = []
self.constraints = {}
self.sparse = sparse
if not name:
prefix = 'input'
name = prefix + '_' + str(K.get_uid(prefix))
self.name = name
if not batch_input_shape:
assert input_shape, 'An Input layer should be passed either a `batch_input_shape` or an `input_shape`.'
batch_input_shape = (None, ) + tuple(input_shape)
else:
batch_input_shape = tuple(batch_input_shape)
if not dtype:
dtype = K.floatx()
self.batch_input_shape = batch_input_shape
self.dtype = dtype
input_tensor = K.placeholder(shape=batch_input_shape,
dtype=dtype,
sparse=self.sparse,
name=self.name)
input_tensor._uses_learning_phase = False
input_tensor._keras_history = (self, 0, 0)
shape = input_tensor._keras_shape
Node(self,
inbound_layers=[],
node_indices=[],
tensor_indices=[],
input_tensors=[input_tensor],
output_tensors=[input_tensor],
input_masks=[None],
output_masks=[None],
input_shapes=[shape],
output_shapes=[shape])
def __init__(
self,
nr_feats: int,
nr_classes: int,
nr_samples: int,
batch_size: int,
epochs: int,
metric_stop_val: float,
hid_layers_sizes: List[int],
activation: str = "tanh",
name: str = None,
ensemble_id=None,
verbose: bool = True,
watch_stop_val: bool = True,
generation: int = 0,
parent_id: str = "no_parent",
net_id: str = None,
rambo_master_id: str = "MISSING_RAMBO_MASTER_ID",
swap_noise: float = None,
dropout: float = 0.0,
):
self.dropout = dropout
super().__init__(
nr_feats,
nr_classes,
nr_samples,
batch_size,
epochs,
metric_stop_val,
hid_layers_sizes,
activation,
name,
ensemble_id,
verbose,
watch_stop_val,
generation,
parent_id,
net_id,
rambo_master_id
)
self.nr_feats = nr_feats
self.swap_noise = swap_noise
self.is_boost_net = False
self.run_fit_count = 0
if not name:
prefix = "rambo_net_"
name = prefix + str(backend.get_uid(prefix))
self.name = name
self.model_path = "models/" + rambo_master_id + "/" + self.net_id + "_keras_model.h5"
def __init__(
self,
nr_feats: int,
nr_classes: int,
nr_samples: int,
batch_size: int,
epochs: int,
metric_stop_val: float,
hid_layers_sizes: List[int],
activation: str = "tanh",
name: str = None,
ensemble_id=None,
verbose: bool = True,
watch_stop_val: bool = True,
generation: int = 0,
parent_id: str = "no_parent",
net_id: str = None,
rambo_master_id: str = "MISSING_RAMBO_MASTER_ID"
):
self.model = Sequential()
self.built = False
self.all_feats = nr_feats
self.hid_layers_sizes = hid_layers_sizes
self.nr_classes = nr_classes
self.nr_samples = nr_samples
self.activation = activation
self.batch_size = batch_size
self.epochs = epochs
self.metric_stop_val = metric_stop_val
self.label_enc = LabelEncoder()
self.ensemble_id = ensemble_id
self.verbose = verbose
self.early_stopper = None
self.watch_stop_val = watch_stop_val
self.bag_train_acc = 0.0
self.oob_train_acc = 0.0
self.full_train_acc = 0.0
self.val_acc = 0.0
self.generation = generation
self.parent_id = parent_id
self.time_created = datetime.datetime.now()
self.test_acc = -1.0
if not name:
prefix = "regular_net_"
name = prefix + str(backend.get_uid(prefix))
self.name = name
self.init_model(nr_feats)
self.net_id = net_id
self.model_path = "models/" + rambo_master_id + "/" + self.net_id + "_keras_model.h5"
def __init__(
self,
inputs,
mode='group',
concat_axis=-1,
output_shape=None,
output_mask=None,
arguments=None,
node_indices=None,
tensor_indices=None,
name=None,
version=1,
):
if K.backend() == 'theano' or K.image_dim_ordering() == 'th':
raise RuntimeError(
"Only support tensorflow backend or image ordering")
self.inputs = inputs
self.mode = mode
self.concat_axis = concat_axis
self._output_shape = output_shape
self.node_indices = node_indices
self._output_mask = output_mask
self.arguments = arguments if arguments else {}
# Layer parameters
self.inbound_nodes = []
self.outbound_nodes = []
self.constraints = {}
self._trainable_weights = []
self._non_trainable_weights = []
self.supports_masking = True
self.uses_learning_phase = False
self.input_spec = None
if not name:
prefix = self.__class__.__name__.lower()
name = prefix + '_' + str(K.get_uid(prefix))
self.name = name
if inputs:
# The inputs is a bunch of nodes shares the same input.
if not node_indices:
node_indices = [0 for _ in range(len(inputs))]
self.built = True
# self.add_inbound_node(inputs, node_indices, tensor_indices)
else:
self.built = False
def make_yolo_depthwise_separable_head(x, num_filters, block_id_str=None):
'''6 Conv2D_BN_Leaky layers followed by a Conv2D_linear layer'''
if not block_id_str:
block_id_str = str(K.get_uid())
x = compose(
DarknetConv2D_BN_Leaky(num_filters, (1, 1)),
Depthwise_Separable_Conv2D_BN_Leaky(filters=num_filters * 2,
kernel_size=(3, 3),
block_id_str=block_id_str + '_1'),
DarknetConv2D_BN_Leaky(num_filters, (1, 1)),
Depthwise_Separable_Conv2D_BN_Leaky(filters=num_filters * 2,
kernel_size=(3, 3),
block_id_str=block_id_str + '_2'),
DarknetConv2D_BN_Leaky(num_filters, (1, 1)))(x)
return x
def BlockGroup(
filters,
strides,
num_repeats,
se_ratio: float = 0.25,
bn_epsilon: float = 1e-5,
bn_momentum: float = 0.0,
activation: str = "relu",
survival_probability: float = 0.8,
name=None,
):
"""Create one group of blocks for the ResNet model."""
if name is None:
counter = backend.get_uid("block_group_")
name = f"block_group_{counter}"
def apply(inputs):
# Only the first block per block_group uses projection shortcut and
# strides.
x = BottleneckBlock(
filters=filters,
strides=strides,
use_projection=True,
se_ratio=se_ratio,
bn_epsilon=bn_epsilon,
bn_momentum=bn_momentum,
activation=activation,
survival_probability=survival_probability,
name=name + "_block_0_",
)(inputs)
for i in range(1, num_repeats):
x = BottleneckBlock(
filters=filters,
strides=1,
use_projection=False,
se_ratio=se_ratio,
activation=activation,
bn_epsilon=bn_epsilon,
bn_momentum=bn_momentum,
survival_probability=survival_probability,
name=name + f"_block_{i}_",
)(x)
return x
return apply
def SE(in_filters: int,
se_ratio: float = 0.25,
expand_ratio: int = 1,
name=None):
"""Squeeze and Excitation block."""
bn_axis = 3 if backend.image_data_format() == "channels_last" else 1
if name is None:
counter = backend.get_uid("se_")
name = f"se_{counter}"
def apply(inputs):
x = layers.GlobalAveragePooling2D(name=name + "_se_squeeze")(inputs)
if bn_axis == 1:
se_shape = (x.shape[-1], 1, 1)
else:
se_shape = (1, 1, x.shape[-1])
x = layers.Reshape(se_shape, name=name + "_se_reshape")(x)
num_reduced_filters = max(1, int(in_filters * 4 * se_ratio))
x = layers.Conv2D(
filters=num_reduced_filters,
kernel_size=[1, 1],
strides=[1, 1],
kernel_initializer=CONV_KERNEL_INITIALIZER,
padding="same",
use_bias=True,
activation="relu",
name=name + "_se_reduce",
)(x)
x = layers.Conv2D(
filters=4 * in_filters *
expand_ratio, # Expand ratio is 1 by default
kernel_size=[1, 1],
strides=[1, 1],
kernel_initializer=CONV_KERNEL_INITIALIZER,
padding="same",
use_bias=True,
activation="sigmoid",
name=name + "_se_expand",
)(x)
return layers.multiply([inputs, x], name=name + "_se_excite")
return apply
def create_input_layer(self, batch_input_shape,
input_dtype=None, name=None):
if not name:
prefix = self.__class__.__name__.lower() + '_input_'
name = prefix + str(K.get_uid(prefix))
if not input_dtype:
input_dtype = K.floatx()
self.batch_input_shape = batch_input_shape
self.input_dtype = input_dtype
# instantiate the input layer
x = SparseInput(batch_shape=batch_input_shape,
dtype=input_dtype, name=name)
# this will build the current layer
# and create the node connecting the current layer
# to the input layer we just created.
self(x)
def __init__(self,
size,
initializer='glorot_uniform',
regularizer=None,
name=None,
**kwargs):
self.size = tuple(size)
self.initializer = initializers.get(initializer)
self.regularizer = regularizers.get(regularizer)
if not name:
prefix = 'shared_weight'
name = prefix + '_' + str(K.get_uid(prefix))
Layer.__init__(self, name=name, **kwargs)
with K.name_scope(self.name):
self.kernel = self.add_weight(shape=self.size,
initializer=self.initializer,
name='kernel',
regularizer=self.regularizer)
self.trainable = True
self.built = True
# self.sparse = sparse
input_tensor = self.kernel * 1.0
self.is_placeholder = False
input_tensor._keras_shape = self.size
input_tensor._uses_learning_phase = False
input_tensor._keras_history = (self, 0, 0)
Node(self,
inbound_layers=[],
node_indices=[],
tensor_indices=[],
input_tensors=[input_tensor],
output_tensors=[input_tensor],
input_masks=[None],
output_masks=[None],
input_shapes=[self.size],
output_shapes=[self.size])
def __init__(self,
shape,
my_initializer='RandomNormal',
name=None,
mult=1.0,
**kwargs):
self.shape = [1, *shape]
self.my_initializer = my_initializer
self.mult = mult
if not name:
prefix = 'param'
name = '%s_%d' % (prefix, K.get_uid(prefix))
Layer.__init__(self, name=name, **kwargs)
# Create a trainable weight variable for this layer.
with K.name_scope(self.name):
self.kernel = self.add_weight(name='kernel',
shape=self.shape,
initializer=self.my_initializer,
trainable=True)
# prepare output tensor, which is essentially the kernel.
output_tensor = self.kernel * self.mult
output_tensor._keras_shape = self.shape
output_tensor._uses_learning_phase = False
output_tensor._keras_history = (self, 0, 0)
output_tensor._batch_input_shape = self.shape
self.trainable = True
self.built = True
self.is_placeholder = False
# create new node
Node(self,
inbound_layers=[],
node_indices=[],
tensor_indices=[],
input_tensors=[],
output_tensors=[output_tensor],
input_masks=[],
output_masks=[None],
input_shapes=[],
output_shapes=[self.shape])
def __init__(self, layers=None, mode='sum', concat_axis=-1,
dot_axes=-1, output_shape=None, output_mask=None,
node_indices=None, tensor_indices=None, name=None):
self.layers = layers
self.mode = mode
self.concat_axis = concat_axis
self.dot_axes = dot_axes
if type(self.dot_axes) == int:
self.dot_axes = [self.dot_axes, ] * 2
self._output_shape = output_shape
self.node_indices = node_indices
self._output_mask = output_mask
# layer parameters
self.inbound_nodes = []
self.outbound_nodes = []
self.constraints = {}
self.regularizers = []
self.trainable_weights = []
self.non_trainable_weights = []
self.supports_masking = True
self.uses_learning_phase = False
self.input_spec = None # compatible with whatever
if not name:
prefix = self.__class__.__name__.lower()
name = prefix + '_' + str(K.get_uid(prefix))
self.name = name
if layers:
# this exists for backwards compatibility.
# equivalent to:
# merge = Merge(layers=None)
# output = merge([input_tensor_1, input_tensor_2])
if not node_indices:
# by default we connect to
# the 1st output stream in the input layer
node_indices = [0 for _ in range(len(layers))]
self._arguments_validation(layers, mode,
concat_axis, dot_axes,
node_indices, tensor_indices)
self.built = True
self.add_inbound_node(layers, node_indices, tensor_indices)
else:
self.built = False
def Depthwise_Separable_Conv2D_BN_Leaky(filters,
kernel_size=(3, 3),
block_id_str=None):
"""Depthwise Separable Convolution2D."""
if not block_id_str:
block_id_str = str(K.get_uid())
return compose(
DepthwiseConv2D(kernel_size,
padding='same',
name='conv_dw_' + block_id_str),
BatchNormalization(name='conv_dw_%s_bn' % block_id_str),
LeakyReLU(alpha=0.1, name='conv_dw_%s_leaky_relu' % block_id_str),
Conv2D(filters, (1, 1),
padding='same',
use_bias=False,
strides=(1, 1),
name='conv_pw_%s' % block_id_str),
BatchNormalization(name='conv_pw_%s_bn' % block_id_str),
LeakyReLU(alpha=0.1, name='conv_pw_%s_leaky_relu' % block_id_str))
def Conv2DFixedPadding(filters, kernel_size, strides, name=None):
"""Conv2D block with fixed padding."""
if name is None:
counter = backend.get_uid("conv_")
name = f"conv_{counter}"
def apply(inputs):
if strides > 1:
inputs = fixed_padding(inputs, kernel_size)
return layers.Conv2D(
filters=filters,
kernel_size=kernel_size,
strides=strides,
padding="same" if strides == 1 else "valid",
use_bias=False,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name,
)(inputs)
return apply
def __init__(self, Auto1, Auto2, layers=None, name=None):
super(MDL_CW, self).__init__()
self.layers = [] # Stack of layers.
self.model = None # Internal Model instance.
self.inputs = [] # List of input tensors
self.outputs = [] # List of length 1: the output tensor (unique).
self._trainable = True
self._initial_weights = None
# Model attributes.
self.inbound_nodes = []
self.outbound_nodes = []
self.built = False
self.pretrain = False
self.Auto1 = Auto1
self.Auto2 = Auto2
self.Unsupervised_train = False
# Set model name.
if not name:
prefix = 'sequential_'
name = prefix + str(K.get_uid(prefix))
self.name = name
def Head(num_classes=1000, name=None):
"""Implementation of classification head of RegNet.
Args:
num_classes: number of classes for Dense layer
name: name prefix
Returns:
Classification head function.
"""
if name is None:
name = str(backend.get_uid("head"))
def apply(x):
x = layers.GlobalAveragePooling2D(name=name + "_head_gap")(x)
x = layers.LayerNormalization(epsilon=1e-6,
name=name + "_head_layernorm")(x)
x = layers.Dense(num_classes, name=name + "_head_dense")(x)
return x
return apply
def __init__(self, name=None, **kwargs):
if not name:
prefix = 'optional_input_placeholder'
name = prefix + '_' + str(K.get_uid(prefix))
kwargs['batch_input_shape'] = (2, )
super(_OptionalInputPlaceHolder, self).__init__(**kwargs)
self.tensor = K.zeros(shape=(2, ))
self.tensor._keras_shape = (2, )
self.tensor._uses_learning_phase = False
self.tensor._keras_history = (self, 0, 0)
Node(self,
inbound_layers=[],
node_indices=[],
tensor_indices=[],
input_tensors=[],
output_tensors=[self.tensor],
input_masks=[None],
output_masks=[None],
input_shapes=[],
output_shapes=[(2, )])
self.build((2, ))
def create_input_layer(self,
batch_input_shape,
input_dtype=None,
name=None):
if not name:
prefix = self.__class__.__name__.lower() + '_input_'
name = prefix + str(K.get_uid(prefix))
if not input_dtype:
input_dtype = K.floatx()
self.batch_input_shape = batch_input_shape
self.input_dtype = input_dtype
# instantiate the input layer
x = SparseInput(batch_shape=batch_input_shape,
dtype=input_dtype,
name=name)
# this will build the current layer
# and create the node connecting the current layer
# to the input layer we just created.
self(x)
def __init__(
self,
input_shape=None,
batch_size=None,
dtype=None,
input_tensor=None,
sparse=None,
name=None,
ragged=None,
type_spec=None,
**kwargs,
):
self._init_input_shape = input_shape
self._init_batch_size = batch_size
self._init_dtype = dtype
self._init_sparse = sparse
self._init_ragged = ragged
self._init_type_spec = type_spec
strategy = tf.distribute.get_strategy()
if (strategy and batch_size is not None
and distributed_training_utils.global_batch_size_supported(
strategy)):
if batch_size % strategy.num_replicas_in_sync != 0:
raise ValueError(
"The `batch_size` argument ({}) must be divisible by "
"the number of replicas ({})".format(
batch_size, strategy.num_replicas_in_sync))
batch_size = batch_size // strategy.num_replicas_in_sync
if "batch_input_shape" in kwargs:
batch_input_shape = kwargs.pop("batch_input_shape")
if input_shape and batch_input_shape:
raise ValueError("Only provide the input_shape OR "
"batch_input_shape argument to "
"InputLayer, not both at the same time.")
# Set the input shape and batch size from the batch_input_shape.
# Note that batch_input_shape can be None (unknown rank) or [] (scalar),
# in which case the batch size must be None.
if batch_input_shape:
batch_size = batch_input_shape[0]
input_shape = batch_input_shape[1:]
if kwargs:
raise ValueError(
f"Unrecognized keyword arguments: {list(kwargs.keys())}")
if sparse and ragged:
raise ValueError(
"Cannot set both sparse and ragged to True in a Keras input.")
if not name:
prefix = "input"
name = prefix + "_" + str(backend.get_uid(prefix))
if not dtype:
if input_tensor is None:
dtype = backend.floatx()
else:
dtype = backend.dtype(input_tensor)
elif input_tensor is not None and input_tensor.dtype != dtype:
raise ValueError(
"`input_tensor.dtype` differs from `dtype`. Received: "
f"input_tensor.dtype={input_tensor.dtype} "
f"but expected dtype={dtype}")
super().__init__(dtype=dtype, name=name)
self.built = True
self.sparse = True if sparse else False
self.ragged = True if ragged else False
self.batch_size = batch_size
self.supports_masking = True
if isinstance(input_shape, tf.TensorShape):
input_shape = tuple(input_shape.as_list())
elif isinstance(input_shape, int):
input_shape = (input_shape, )
if type_spec is not None:
args_that_must_be_none = [
("(input_)shape", self._init_input_shape),
("batch_size", self._init_batch_size),
("dtype", self._init_dtype),
("input_tensor", input_tensor),
("sparse", self._init_sparse),
("ragged", self._init_ragged),
]
for arg_name, arg in args_that_must_be_none:
_assert_other_arg_none(arg_name, arg)
if not tf.compat.v1.executing_eagerly_outside_functions():
raise ValueError(
"Creating Keras inputs from a type_spec is only "
"supported when eager execution is enabled.")
input_tensor = keras_tensor.keras_tensor_from_type_spec(type_spec)
if isinstance(input_tensor, keras_tensor.SparseKerasTensor):
self.sparse = True
if isinstance(input_tensor, keras_tensor.RaggedKerasTensor):
self.ragged = True
self.is_placeholder = True
try:
self._batch_input_shape = tuple(input_tensor.shape.as_list())
except ValueError:
# If the shape cannot be represented as a tuple (e.g. unknown rank)
self._batch_input_shape = None
elif input_tensor is None:
if input_shape is not None:
batch_input_shape = (batch_size, ) + tuple(input_shape)
else:
batch_input_shape = None
graph = backend.get_graph()
with graph.as_default():
input_tensor = backend.placeholder(
shape=batch_input_shape,
dtype=dtype,
name=self.name,
sparse=sparse,
ragged=ragged,
)
self.is_placeholder = True
self._batch_input_shape = batch_input_shape
else:
if tf.compat.v1.executing_eagerly_outside_functions():
if not isinstance(input_tensor, keras_tensor.KerasTensor):
input_tensor = keras_tensor.keras_tensor_from_tensor(
input_tensor)
else:
if not tf_utils.is_symbolic_tensor(input_tensor):
raise ValueError(
"You should not pass an EagerTensor to `Input`. "
"For example, instead of creating an "
"`InputLayer`, you should instantiate your model "
"and directly call it on your input.")
self.is_placeholder = False
try:
self._batch_input_shape = tuple(input_tensor.shape.as_list())
except ValueError:
# If the shape cannot be represented as a tuple (e.g. unknown rank)
self._batch_input_shape = None
# Create an input node.
input_tensor._keras_mask = None
node_module.Node(layer=self, outputs=input_tensor)
# Store type spec
if isinstance(input_tensor, keras_tensor.KerasTensor) or (
tf_utils.is_extension_type(input_tensor)):
self._type_spec = (input_tensor._type_spec) # pylint: disable=protected-access
else:
self._type_spec = tf.TensorSpec(
shape=input_tensor.shape,
dtype=input_tensor.dtype,
name=self.name,
)
def _get_unique_name(name, prefix=None):
if prefix is not None:
name = prefix + '_' + name
name += '_' + str(K.get_uid(name))
return name
def _unique(name):
"""
Return a unique name string.
"""
return name + '_' + str(K.get_uid(name))
def FusedMBConvBlock(
input_filters: int,
output_filters: int,
expand_ratio=1,
kernel_size=3,
strides=1,
se_ratio=0.0,
bn_momentum=0.9,
activation="swish",
survival_probability: float = 0.8,
name=None,
):
"""Fused MBConv Block: Fusing the proj conv1x1 and depthwise_conv into a conv2d."""
bn_axis = 3 if backend.image_data_format() == "channels_last" else 1
if name is None:
name = backend.get_uid("block0")
def apply(inputs):
filters = input_filters * expand_ratio
if expand_ratio != 1:
x = layers.Conv2D(
filters,
kernel_size=kernel_size,
strides=strides,
kernel_initializer=CONV_KERNEL_INITIALIZER,
data_format="channels_last",
padding="same",
use_bias=False,
name=name + "expand_conv",
)(inputs)
x = layers.BatchNormalization(axis=bn_axis,
momentum=bn_momentum,
name=name + "expand_bn")(x)
x = layers.Activation(activation=activation,
name=name + "expand_activation")(x)
else:
x = inputs
# Squeeze and excite
if 0 < se_ratio <= 1:
filters_se = max(1, int(input_filters * 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(
output_filters,
kernel_size=1 if expand_ratio != 1 else kernel_size,
strides=1 if expand_ratio != 1 else strides,
kernel_initializer=CONV_KERNEL_INITIALIZER,
padding="same",
use_bias=False,
name=name + "project_conv",
)(x)
x = layers.BatchNormalization(axis=bn_axis,
momentum=bn_momentum,
name=name + "project_bn")(x)
if expand_ratio == 1:
x = layers.Activation(activation=activation,
name=name + "project_activation")(x)
# Residual:
if strides == 1 and input_filters == output_filters:
if survival_probability:
x = layers.Dropout(
survival_probability,
noise_shape=(None, 1, 1, 1),
name=name + "drop",
)(x)
x = layers.add([x, inputs], name=name + "add")
return x
return apply
def BottleneckBlock(
filters: int,
strides: int,
use_projection: bool,
bn_momentum: float = 0.0,
bn_epsilon: float = 1e-5,
activation: str = "relu",
se_ratio: float = 0.25,
survival_probability: float = 0.8,
name=None,
):
"""Bottleneck block variant for residual networks with BN."""
if name is None:
counter = backend.get_uid("block_0_")
name = f"block_0_{counter}"
def apply(inputs):
bn_axis = 3 if backend.image_data_format() == "channels_last" else 1
shortcut = inputs
if use_projection:
filters_out = filters * 4
if strides == 2:
shortcut = layers.AveragePooling2D(
pool_size=(2, 2),
strides=(2, 2),
padding="same",
name=name + "_projection_pooling",
)(inputs)
shortcut = Conv2DFixedPadding(
filters=filters_out,
kernel_size=1,
strides=1,
name=name + "_projection_conv",
)(shortcut)
else:
shortcut = Conv2DFixedPadding(
filters=filters_out,
kernel_size=1,
strides=strides,
name=name + "_projection_conv",
)(inputs)
shortcut = layers.BatchNormalization(
axis=bn_axis,
momentum=bn_momentum,
epsilon=bn_epsilon,
name=name + "_projection_batch_norm",
)(shortcut)
# First conv layer:
x = Conv2DFixedPadding(filters=filters,
kernel_size=1,
strides=1,
name=name + "_conv_1")(inputs)
x = layers.BatchNormalization(
axis=bn_axis,
momentum=bn_momentum,
epsilon=bn_epsilon,
name=name + "batch_norm_1",
)(x)
x = layers.Activation(activation, name=name + "_act_1")(x)
# Second conv layer:
x = Conv2DFixedPadding(
filters=filters,
kernel_size=3,
strides=strides,
name=name + "_conv_2",
)(x)
x = layers.BatchNormalization(
axis=bn_axis,
momentum=bn_momentum,
epsilon=bn_epsilon,
name=name + "_batch_norm_2",
)(x)
x = layers.Activation(activation, name=name + "_act_2")(x)
# Third conv layer:
x = Conv2DFixedPadding(filters=filters * 4,
kernel_size=1,
strides=1,
name=name + "_conv_3")(x)
x = layers.BatchNormalization(
axis=bn_axis,
momentum=bn_momentum,
epsilon=bn_epsilon,
name=name + "_batch_norm_3",
)(x)
if 0 < se_ratio < 1:
x = SE(filters, se_ratio=se_ratio, name=name + "_se")(x)
# Drop connect
if survival_probability:
x = layers.Dropout(
survival_probability,
noise_shape=(None, 1, 1, 1),
name=name + "_drop",
)(x)
x = layers.Add()([x, shortcut])
return layers.Activation(activation, name=name + "_output_act")(x)
return apply
