def create(input_width, input_height, input_channels=1):
def instantiate(**_):
return DoubleNatureCnn(input_width=input_width,
input_height=input_height,
input_channels=input_channels)
return ModelFactory.generic(instantiate)
def create(input_width, input_height, input_channels=1, output_dim=512):
def instantiate(**_):
return NatureCnnTwoTower(
input_width=input_width, input_height=input_height, input_channels=input_channels,
output_dim=output_dim
)
return ModelFactory.generic(instantiate)
def create(input_width, input_height, input_channels=1, cnn_output_dim=512, hidden_units=128):
def instantiate(**_):
return NatureCnnLstmBackbone(
input_width=input_width, input_height=input_height, input_channels=input_channels,
cnn_output_dim=cnn_output_dim, hidden_units=hidden_units
)
return ModelFactory.generic(instantiate)
def create(input_block: LinearBackboneModel, hidden_layers: typing.List[int], output_dim: int, dropout=0.0):
""" Vel creation function """
def instantiate(**_):
return MultilayerSequenceLSTM(
input_block, hidden_layers, output_dim, dropout=dropout
)
return ModelFactory.generic(instantiate)
def create(input_length, hidden_layers, activation='tanh', normalization=None):
def instantiate(**_):
return MLP(input_length=input_length,
hidden_layers=hidden_layers,
activation=activation,
normalization=normalization)
return ModelFactory.generic(instantiate)
def create(input_length,
layers=2,
hidden_units=64,
activation='tanh',
layer_norm=True):
def instantiate(**_):
return MLP(input_length=input_length,
layers=layers,
hidden_units=hidden_units,
activation=activation,
layer_norm=layer_norm)
return ModelFactory.generic(instantiate)
def create(blocks, mode='basic', inplanes=64, cardinality=4, image_features=64, divisor=4, num_classes=1000):
"""
Create a ResNetV1 model
"""
block_dict = {
# 'basic': BasicBlock,
'bottleneck': ResNeXtBottleneck
}
def instantiate(**_):
return ResNeXt(block_dict[mode], blocks, inplanes=inplanes, image_features=image_features, cardinality=cardinality, divisor=divisor, num_classes=num_classes)
return ModelFactory.generic(instantiate)
def create(blocks, mode='basic', inplanes=16, divisor=4, num_classes=1000):
"""
Create a ResNetV1 model
"""
block_dict = {'basic': BasicBlock, 'bottleneck': Bottleneck}
def instantiate(**_):
return ResNetV2(block_dict[mode],
blocks,
inplanes=inplanes,
divisor=divisor,
num_classes=num_classes)
return ModelFactory.generic(instantiate)
def create(input_block: LinearBackboneModel,
output_dim: int,
rnn_layers: typing.List[int],
rnn_dropout: float = 0.0,
bidirectional: bool = False,
linear_layers: typing.List[int] = None,
linear_dropout: float = 0.0):
""" Vel creation function """
if linear_layers is None:
linear_layers = []
def instantiate(**_):
return MultilayerSequenceClassificationGRU(
input_block=input_block,
output_dim=output_dim,
rnn_layers=rnn_layers,
rnn_dropout=rnn_dropout,
bidirectional=bidirectional,
linear_layers=linear_layers,
linear_dropout=linear_dropout)
return ModelFactory.generic(instantiate)
def create(fc_layers=None, dropout=None, pretrained=True):
""" Create a Resnet-34 model with a custom head """
def instantiate(**_):
return Resnet34(fc_layers, dropout, pretrained)
return ModelFactory.generic(instantiate)
def create(img_rows, img_cols, img_channels, num_classes):
""" Create the model matching specified image dimensions """
def instantiate(**_):
return Net(img_rows, img_cols, img_channels, num_classes)
return ModelFactory.generic(instantiate)
