2D Convolutional Autoencoder

[dreamer2368]

Neural network module

lasdi.networks module provides various NN modules that can be combined into encoder/decoder.

• MultiLayerPerceptron class is moved from previous lasdi.latent_space module.
• CNN2D class is newly added for 2D convolutional neural networks.
  • channels/kernel sizes for Conv2d are automatically determined from given architecture.
    • class methods compute_kernel_size, compute_input_layer_size, compute_output_layer_size compute the resulting kernel/input layer/output layer size based on given architecture.
  • NOTE torch.nn.Conv2d only takes 3d-/4d- tensors, while training data can have more flexible shapes.
    • Before training, the training data shape must be passed via CNN2D.set_data_shape, so that the data can be properly reshape to/from 3d-/4d tensors.
• Fully documented according to numpy style.

LatentSpace base class

• Latent space classes are moved to submodule lasdi.latent_space.
• Base class LatentSpace provides the basic skeleton of latent space classes.
• Documentation will be done in future PRs.

Conv2DAutoencoder class

• Derived class of LatentSpace
• Encoder: CNN2D - MultiLayerPerceptron
• Decoder: MultiLayerPerceptron - CNN2D
• NOTE: Before training, the training data shape must be passed via Conv2DAutoencoder.set_batch_shape.

Miscellaneous change

• CI workflow for unit tests
• CNN2D unit tests
• BayesianGLaSDI save loss history and load it from restart file. Training appends from the loaded loss history.