Developer Guide Networks

(work in progress)

Scope

monai.networks consists of nets, blocks and layers sub-modules that correspond to

• network model implementations,
• reusable building blocks,
• essential differentiable component, respectively.

blocks is implemented as a pytorch module with one or multiple layers. nets is implemented as a pytorch module with one or multiple layers and blocks.

Features

• compositional design to create complex networks
• 1d/2d/3d with the factories https://github.com/Project-MONAI/MONAI/blob/dev/monai/networks/layers/factories.py

dimension=3name=Conv.CONVTRANSconv=Conv[name, dimension]

• flexible layer configuration https://github.com/Project-MONAI/MONAI/blob/dev/monai/networks/layers/utils.py

frommonai.networks.layersimportget_act_layers_layer=get_act_layer(name="swish")
p_layer=get_act_layer(name=("prelu", {"num_parameters": 1, "init": 0.25}))

• support of torchscript

Tests

• basic input/output shape and datatype tests, e.g. https://github.com/Project-MONAI/MONAI/blob/dev/tests/test_se_block.py
• torch script tests, e.g. SE block test case
• network integration tests https://github.com/Project-MONAI/MONAI/blob/dev/tests/test_network_consistency.py