aixd.mlmodel

This package contains the machine-learning model implementation.

Data Loading

DataModule

Data module for the ML model.

Architecture

Models

CondAEModel	Class representing a Conditional Autoencoder model.
CondVAEModel	Class representing a Conditional Variational Autoencoder model.
InverseModel	Wrapper class for taking a Cond(V)AEModel and fine-tuning the decoder using the encoder as surrogate model.
FreezeEncoder	A PyTorch Lightning callback class that extends the BaseFinetuning class and freezes the encoder of the CondAEModel before training.

Blocks

ResBlockFC	Module implementing a fully-connected residual block.
ResBlockConv	Module implementing a convolutional residual block.
ResBlock1D	Module implementing a 1D convolutional residual block.
ResBlock2D	Module implementing a 2D convolutional residual block.
ResBlock3D	Module implementing a 3D convolutional residual block.
SelfAttn	Module implementing a self-attention layer.
SelfAttn1D	Module implementing a 1D self-attention layer.
SelfAttn2D	Module implementing a 2D self-attention layer.
SelfAttn3D	Module implementing a 3D self-attention layer.
Activation	Module wrapping around an activation function.

Decoders/Encoders

Decoder	Decoder module of the conditional (variational) autoencoder.
Encoder	Encoder module of a conditional variational autoencoder.
VEncoder	Encoder module of a conditional variational autoencoder with additional parameters for the latent distribution.

Heads

InHeadFC	Fully-Connected Feed-Forward Network for encoding an unstrucured, 1D feature.
OutHeadFC	Fully-Connected Feed-Forward Network for decoding an unstrucured, 1D feature.
InHeadConv	Convolutional Network for encoding spatially strucured data in 1, 2 or 3 dimensions.
OutHeadConv	Convolutional Network for decoding spatially strucured data in 1, 2 or 3 dimensions.
InHeadConv1D	Convolutional Network for encoding spatially strucured (temporal) data in 1 dimension.
OutHeadConv1D	Convolutional Network for decoding spatially strucured (temporal) data in 1 dimension.
InHeadConv2D	Convolutional Network for encoding spatially strucured (image-like) data in 2 dimensions.
OutHeadConv2D	Convolutional Network for decoding spatially strucured (image-like) data in 2 dimensions.
InHeadConv3D	Convolutional Network for encoding spatially strucured (MRI- or video-like) data in 3 dimensions.
OutHeadConv3D	Convolutional Network for decoding spatially strucured (MRI- or video-like) data in 3 dimensions.

Losses

LossStd	Loss for the standard deviation of the values.
MGEloss	Mean gradient error.

Generation

Generator	Initialize a Generator instance.
GeneratorSampler	The Generator class needs to sampled from the outputML, in order to provide these values as entry to the decoder.