beyondml.pt package

Subpackages

• beyondml.pt.layers package
  • Submodules
  • beyondml.pt.layers.Conv2D module
    • Conv2D
      • Conv2D.forward()
  • beyondml.pt.layers.Conv3D module
    • Conv3D
      • Conv3D.forward()
  • beyondml.pt.layers.Dense module
    • Dense
      • Dense.forward()
  • beyondml.pt.layers.FilterLayer module
    • FilterLayer
      • FilterLayer.forward()
      • FilterLayer.is_on
      • FilterLayer.turn_off()
      • FilterLayer.turn_on()
  • beyondml.pt.layers.MaskedConv2D module
    • MaskedConv2D
      • MaskedConv2D.forward()
      • MaskedConv2D.in_channels
      • MaskedConv2D.kernel_size
      • MaskedConv2D.out_channels
      • MaskedConv2D.prune()
  • beyondml.pt.layers.MaskedConv3D module
    • MaskedConv3D
      • MaskedConv3D.forward()
      • MaskedConv3D.in_channels
      • MaskedConv3D.kernel_size
      • MaskedConv3D.out_channels
      • MaskedConv3D.prune()
  • beyondml.pt.layers.MaskedDense module
    • MaskedDense
      • MaskedDense.forward()
      • MaskedDense.prune()
  • beyondml.pt.layers.MaskedMultiHeadAttention module
    • MaskedMultiHeadAttention
      • MaskedMultiHeadAttention.forward()
      • MaskedMultiHeadAttention.prune()
  • beyondml.pt.layers.MaskedTransformerDecoderLayer module
    • MaskedTransformerDecoderLayer
      • MaskedTransformerDecoderLayer.forward()
      • MaskedTransformerDecoderLayer.prune()
  • beyondml.pt.layers.MaskedTransformerEncoderLayer module
    • MaskedTransformerEncoderLayer
      • MaskedTransformerEncoderLayer.forward()
      • MaskedTransformerEncoderLayer.prune()
  • beyondml.pt.layers.MultiConv2D module
    • MultiConv2D
      • MultiConv2D.forward()
  • beyondml.pt.layers.MultiConv3D module
    • MultiConv3D
      • MultiConv3D.forward()
  • beyondml.pt.layers.MultiDense module
    • MultiDense
      • MultiDense.forward()
  • beyondml.pt.layers.MultiMaskedConv2D module
    • MultiMaskedConv2D
      • MultiMaskedConv2D.forward()
      • MultiMaskedConv2D.in_channels
      • MultiMaskedConv2D.kernel_size
      • MultiMaskedConv2D.out_channels
      • MultiMaskedConv2D.prune()
  • beyondml.pt.layers.MultiMaskedConv3D module
    • MultiMaskedConv3D
      • MultiMaskedConv3D.forward()
      • MultiMaskedConv3D.in_channels
      • MultiMaskedConv3D.kernel_size
      • MultiMaskedConv3D.out_channels
      • MultiMaskedConv3D.prune()
  • beyondml.pt.layers.MultiMaskedDense module
    • MultiMaskedDense
      • MultiMaskedDense.forward()
      • MultiMaskedDense.prune()
  • beyondml.pt.layers.MultiMaxPool2D module
    • MultiMaxPool2D
      • MultiMaxPool2D.forward()
  • beyondml.pt.layers.MultiMaxPool3D module
    • MultiMaxPool3D
      • MultiMaxPool3D.forward()
  • beyondml.pt.layers.MultitaskNormalization module
    • MultitaskNormalization
      • MultitaskNormalization.forward()
  • beyondml.pt.layers.SelectorLayer module
    • SelectorLayer
      • SelectorLayer.forward()
      • SelectorLayer.sel_index
  • beyondml.pt.layers.SparseConv2D module
    • SparseConv2D
      • SparseConv2D.forward()
  • beyondml.pt.layers.SparseConv3D module
    • SparseConv3D
      • SparseConv3D.forward()
  • beyondml.pt.layers.SparseDense module
    • SparseDense
      • SparseDense.forward()
  • beyondml.pt.layers.SparseMultiConv2D module
    • SparseMultiConv2D
      • SparseMultiConv2D.forward()
  • beyondml.pt.layers.SparseMultiConv3D module
    • SparseMultiConv3D
      • SparseMultiConv3D.forward()
  • beyondml.pt.layers.SparseMultiDense module
    • SparseMultiDense
      • SparseMultiDense.forward()
  • Module contents
• beyondml.pt.utils package
  • Submodules
  • beyondml.pt.utils.utils module
    • prune_model()
  • Module contents

Module contents

## PyTorch compatibility for building MANN models

The beyondml.pt subpackage contains layers and utilities for creating and pruning models using [PyTorch](https://pytorch.org). The package contains two subpackages, the beyondml.pt.layers package, and the beyondml.pt.utils package.

Within the layers package, there is current functionality for the the following layers: - beyondml.pt.layers.Conv2D - beyondml.pt.layers.Dense - beyondml.pt.layers.FilterLayer - beyondml.pt.layers.MaskedConv2D - beyondml.pt.layers.MaskedDense - beyondml.pt.layers.MultiConv2D - beyondml.pt.layers.MultiDense - beyondml.pt.layers.MultiMaskedConv2D - beyondml.pt.layers.MultiMaskedDense - beyondml.pt.layers.SelectorLayer - beyondml.pt.layers.SparseConv2D - beyondml.pt.layers.SparseDense - beyondml.pt.layers.SparseMultiConv2D - beyondml.pt.layers.SparseMultiDense

Within the beyondml.pt.utils package, there is currently only one function, the prune_model function. Because of the openness of developing with PyTorch in comparison to TensorFlow, there is far less functionality that can be supplied directly via BeyondML. Instead, for converting models from training to inference, the user is left to devise the best way to do so by building his or her own classes.

### Best Practices for Pruning In order to use the utils.prune_model function, the model itself must have a .layers property. This property is used to determine which layers can be pruned. Only layers which support pruning and which are included in the `.layers` property are pruned, meaning the user can determine which exact layers in the model he or she wants pruned. Alternatively, the user can create their own pruning function or method on the class itself and prune that way, utilizing each of the .prune() methods of the layers provided.