torch-lego

Build pytorch modules using yaml description files. Yaml can be used to keep track of model architecture using tools such as Mlflow. This library supports convolution (1d, 2d and 3d) modules, transpose convolution (1d,2d,3d) modules and linear modules. Average and max pooling layers are also supported as well as fully connected layers. Batch/Layer/Instance norm, dropout and all built-in PyTorch activations are available. See the examples in the test directory for more detail.

Example

Deep convolutional classifier

The example given below is a simple MNIST classifier. LeakyReLU are used as activations. The negative slope value is not specfied so the default PyTorch value will be used. The same applies for all non-specified parameters. Max pooling is used and the input size is automatically computed using parameters from the preceeding convolution layer.

architecture:
  encoder:
    - conv:
        in_channels: 1
        out_channels: 36
        kernel_size: [3,3]
        stride: 1
        padding: 2
      mpool:
        kernel_size: [2,2]
      act:
        type: 'LeakyReLU'
        params:
    - conv:
        in_channels: 36
        out_channels: 100
        kernel_size: [3,3]
        stride: 1
        padding: 2
      mpool:
        kernel_size: [ 2,2 ]
      act:
        type: 'LeakyReLU'
        params:
      drpt: 0.5
  classifier:
    - in_features: 6400
    - out_features: 200
      act:
        type: 'ReLU'
        params:
      drpt: 0.4
    - out_features: 100
      act:
        type: 'ReLU'
        params:
      drpt: 0.4
    - out_features: 10
      act:
        type:
        params:
      drpt: 0.2

Installation and test

This is a python package that can be installed using pip after cloning the repository:

pip install -e torch_lego

All tests can be run by executing the all.sh script in the tests directory.