Step 3: Compress the model

This section provides NetsPresso Compressor, which enhances computational efficiency. NetsPresso Compressor provides various compression method, including structured pruning and filter decomposition.

Range of Support

The range of compression methods in the Python package can be found in the API documentation (link), while those applied in the GUI are listed below:

Method	GUI	Python
Automatic Compression	❌	✅
Structured Pruning_Geometric Median	✅	✅
Structured Pruning_L2Norm	✅	✅
Structured Pruning_Structured Neuron-level Pruning (SNP)	❌	✅
Structured Pruning_Nuclear Norm Pruning	❌	✅
Filter Decomposition_Singular Value Decomposition	✅	✅
Filter Decomposition_Tucker Decomposition	✅	✅
Filter Decomposition_CP Decomposition	❌	✅

Compress Configuration

The GUI Compression configuration supports multiple experiments, allowing up to 12 experiments per compression method.

Structured Pruning
- Geometric Median
- L2Norm
Filter Decomposition
- Singular Value Decomposition
- Tucker Decomposition

Once the experiment results are complete, the user will be redirected to the main page, where they can check the experiment status via colored dots.