Compression Operators

condensa.pi.blockprune(module, threshold, block_size, criteria, align=None, parameter='weight')

Prunes blocks of module parameters based on magnitude (inplace).

Parameters

  • module (torch.nn.Module) – PyTorch module.

  • threshold (float) – Magnitude threshold for pruning.

  • block_size (Tuple) – Block size for pruning.

  • criteria (condensa.functional) – Aggregation function for thresholding.

  • align (int) – Alignment of compressed parameters.

  • parameter (str) – Module parameter to prune (default: ‘weight’)

condensa.pi.filter_prune(module, threshold, criteria, align=None, prune_bias=True)

Prunes 3D blocks (filters) of module parameters based on magnitude (inplace).

Parameters

  • module (torch.nn.Module) – PyTorch module.

  • threshold (float) – Magnitude threshold for pruning.

  • criteria (condensa.functional) – Aggregation function for thresholding.

  • align (int) – Alignment of compressed parameters.

  • prune_bias (bool) – Whether to prune corresponding biases.

condensa.pi.neuron_prune(module, threshold, criteria, align=None, prune_bias=True)

Prunes neurons based on magnitude (inplace).

Parameters

  • module (torch.nn.Module) – PyTorch module.

  • threshold (float) – Magnitude threshold for pruning.

  • criteria (condensa.functional) – Aggregation function for thresholding.

  • align (int) – Alignment of compressed parameters.

  • prune_bias (bool) – Whether to prune corresponding biases.

condensa.pi.prune(module, threshold, parameter='weight')

Prunes module parameters based on magnitude (inplace).

Parameters

  • module (torch.nn.Module) – PyTorch module.

  • threshold (float) – Magnitude threshold for pruning.

  • parameter (str) – Module parameter to prune (default: ‘weight’)

condensa.pi.quantize(module, dtype)

Quantizes module to given data type (inplace).

Parameters

  • module (torch.nn.Module) – PyTorch module.

  • dtype – Target data type.

condensa.delta.dequantize(module, dtype)

De-quantizes module to given data type (inplace).

Parameters

  • module (torch.nn.Module) – PyTorch module.

  • dtype – Target data type.