def __init__(
self,
input_shape,
inner_dim,
activation=torch.nn.Sigmoid,
norm=BatchNorm1d,
):
super().__init__()
self.inner_dim = inner_dim
self.norm = norm
self.activation = activation
bz, t, chn = input_shape
self.conv = Sequential(input_shape=input_shape)
self.conv.append(
DepthwiseSeparableConv1d,
out_channels=chn,
kernel_size=1,
stride=1,
)
self.conv.append(self.norm)
self.conv.append(self.activation())
self.avg_pool = AdaptivePool(1)
self.bottleneck = Sequential(
Linear(input_size=input_shape[-1], n_neurons=self.inner_dim),
self.activation(),
Linear(input_size=self.inner_dim, n_neurons=chn),
self.activation(),
)
def __init__(
self,
out_channels,
kernel_size,
num_layers,
inner_dim,
input_shape,
stride=1,
beta=1,
dropout=0.15,
activation=Swish,
se_activation=torch.nn.Sigmoid,
norm=BatchNorm1d,
residual=True,
):
super().__init__()
self.residual = residual
self.Convs = Sequential(input_shape=input_shape)
for i in range(num_layers):
self.Convs.append(
DepthwiseSeparableConv1d,
out_channels,
kernel_size,
stride=stride if i == num_layers - 1 else 1,
)
self.Convs.append(norm)
self.SE = SEmodule(
input_shape=self.Convs.get_output_shape(),
inner_dim=inner_dim,
activation=se_activation,
norm=norm,
)
self.drop = Dropout(dropout)
self.reduced_cov = None
if residual:
self.reduced_cov = Sequential(input_shape=input_shape)
self.reduced_cov.append(
Conv1d,
out_channels,
kernel_size=3,
stride=stride,
)
self.reduced_cov.append(norm)
if isinstance(activation, Swish):
self.activation = activation(beta)
else:
self.activation = activation()
self._reset_params()
def __init__(
self,
num_layers,
out_channels,
input_shape,
kernel_size=3,
stride=1,
dilation=1,
residual=False,
conv_module=Conv2d,
activation=torch.nn.LeakyReLU,
norm=None,
dropout=0.1,
):
super().__init__()
self.convs = Sequential(input_shape=input_shape)
for i in range(num_layers):
self.convs.append(
conv_module,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride if i == num_layers - 1 else 1,
dilation=dilation,
layer_name=f"conv_{i}",
)
if norm is not None:
self.convs.append(norm, layer_name=f"norm_{i}")
self.convs.append(activation(), layer_name=f"act_{i}")
self.convs.append(
torch.nn.Dropout(dropout), layer_name=f"dropout_{i}"
)
self.reduce_conv = None
self.drop = None
if residual:
self.reduce_conv = Sequential(input_shape=input_shape)
self.reduce_conv.append(
conv_module,
out_channels=out_channels,
kernel_size=1,
stride=stride,
layer_name="conv",
)
self.reduce_conv.append(norm, layer_name="norm")
self.drop = torch.nn.Dropout(dropout)