chop.passes.module.transform.quantize

chop.passes.module.transform.quantize#

calculate_avg_bits_module_analysis_pass#

chop.passes.module.analysis.quantize.calculate_avg_bits_module_analysis_pass(module: Module, pass_args: dict = {}) → tuple[source]#

Analyzes the averaged weight bitwidth of a given module. Considering only Linear and Conv2d layers.

Parameters:

  • module (torch.nn.Module) – The module to analyze.

  • pass_args (dict) – Additional arguments for the analysis pass. (default: {})

Returns:

A tuple containing the modified module and a dictionary with the analysis results.

Return type:

tuple(torch.nn.Module, dict)

analysis results is a dictionary with the following keys:

  • ‘average_bitwidth’ (float): The average number of bits per value for weight.

Examples output:

info = {
    # this means on average each weight value is represented by 16 bits
    'average_bitwidth': 16.0}