piscis.networks.conv

piscis.networks.conv#

Attributes#

`ModuleDef`
`ConvBatchAct`

Classes#

`Conv`	Convolutional block with batch norm and activation.
`SqueezeExcite`	Squeeze and excite block.
`MBConv`	Mobile inverted bottleneck convolutional block.
`FusedMBConv`	Fused mobile inverted bottleneck convolutional block.

Module Contents#

piscis.networks.conv.ModuleDef#

class piscis.networks.conv.Conv(in_channels: int, out_channels: int, kernel_size: int | Sequence[int] = 3, stride: int | Sequence[int] = 1, padding: int | Sequence[int] | str | None = None, dilation: int | Sequence[int] = 1, groups: int = 1, bias: bool | None = None, padding_mode: str = 'zeros', conv: ModuleDef = nn.Conv2d, bn: ModuleDef = nn.BatchNorm2d, act: ModuleDef = nn.SiLU, layers: Sequence[str] = ('conv', 'bn', 'act'))#

Bases: torch.nn.Module

Convolutional block with batch norm and activation.

Parameters:

in_channelsint: Number of input channels.
out_channelsint: Number of output channels.
kernel_sizeUnion[int, Sequence[int]]: Size of the convolutional kernel. Default is 3.
strideUnion[int, Sequence[int]]: Stride of the convolution. Default is 1.
paddingOptional[Union[int, Sequence[int], str]], optional: Padding of the convolution. Default is None.
dilationUnion[int, Sequence[int]], optional: Dilation of the convolution. Default is 1.
groupsint, optional: Number of groups of the convolution. Default is 1.
biasOptional[bool], optional: Whether to use bias in the convolution. Default is None.
padding_modestr, optional: Padding mode of the convolution. Default is ‘zeros’.
convModuleDef, optional: Convolution module. Default is nn.Conv2d.
bnModuleDef, optional: Batch norm module. Default is nn.BatchNorm2d.
actModuleDef, optional: Activation function. Default is nn.SiLU.
layersSequence[str], optional: Sequence of layers to apply. Default is (‘conv’, ‘bn’, ‘act’).

conv#

forward(x)#

piscis.networks.conv.ConvBatchAct#

class piscis.networks.conv.SqueezeExcite(in_channels: int, squeeze_channels: int, conv: ModuleDef = nn.Conv2d, act: ModuleDef = nn.SiLU)#

Bases: torch.nn.Module

Squeeze and excite block.

Parameters:

in_channelsint: Number of input channels.
squeeze_channelsint: Number of squeeze channels.
convModuleDef, optional: Convolution module. Default is nn.Conv2d.
actModuleDef, optional: Activation function. Default is nn.SiLU.

reduce#

act#

expand#

sigmoid#

forward(x: torch.Tensor) → torch.Tensor#

class piscis.networks.conv.MBConv(in_channels: int, out_channels: int, expand_ratio: int, kernel_size: int | Sequence[int], stride: int, se_ratio: float, stochastic_depth_prob: float, conv: ModuleDef = nn.Conv2d, bn: ModuleDef = nn.BatchNorm2d, act: ModuleDef = nn.SiLU)#

Bases: torch.nn.Module

Mobile inverted bottleneck convolutional block.

Parameters:

in_channelsint: Number of input channels.
out_channelsint: Number of output channels.
expand_ratioint: Expansion ratio.
kernel_sizeUnion[int, Sequence[int]]: Size of the convolutional kernel.
strideint: Stride of the convolution.
se_ratiofloat: Squeeze and excitation ratio.
stochastic_depth_probfloat: Stochastic depth probability.
convModuleDef, optional: Convolution module. Default is nn.Conv2d.
bnModuleDef, optional: Batch norm module. Default is nn.BatchNorm2d.
actModuleDef, optional: Activation function. Default is nn.SiLU.

block#

use_res_connect#

stochastic_depth#

forward(x: torch.Tensor) → torch.Tensor#

class piscis.networks.conv.FusedMBConv(in_channels: int, out_channels: int, expand_ratio: int, kernel_size: int | Sequence[int], stride: int, se_ratio: float, stochastic_depth_prob: float, conv: ModuleDef = nn.Conv2d, bn: ModuleDef = nn.BatchNorm2d, act: ModuleDef = nn.SiLU)#

Bases: torch.nn.Module

Fused mobile inverted bottleneck convolutional block.

Parameters:

in_channelsint: Number of input channels.
out_channelsint: Number of output channels.
expand_ratioint: Expansion ratio.
kernel_sizeUnion[int, Sequence[int]]: Size of the convolutional kernel.
strideint: Stride of the convolution.
se_ratiofloat: Squeeze and excitation ratio.
stochastic_depth_probfloat: Stochastic depth probability.
convModuleDef, optional: Convolution module. Default is nn.Conv2d.
bnModuleDef, optional: Batch norm module. Default is nn.BatchNorm2d.
actModuleDef, optional: Activation function. Default is nn.SiLU.

block#

use_res_connect#

stochastic_depth#

forward(x: torch.Tensor) → torch.Tensor#