How to get the weights of quantized module?

marka17 · March 4, 2021, 6:52pm

Hi!

Let me have the following module:

m = nn.quantized.dynamic.Linear(20, 30, dtype=torch.qint8)
m = torch.jit.script(m)

How can I get the weights of the m module back?

raghuramank100 · March 8, 2021, 1:07am

Use the _weight_bias() method to get the weights back:

(wt,bias) = m._packed_params._weight_bias()

github.com

pytorch/pytorch/blob/main/torch/nn/quantized/modules/linear.py#L32


      
          # flake8: noqa: F401
          r"""Quantized Modules.
          
          This file is in the process of migration to `torch/ao/nn/quantized`, and
          is kept here for compatibility while the migration process is ongoing.
          If you are adding a new entry/functionality, please, add it to the
          appropriate file under the `torch/ao/nn/quantized/modules`,
          while adding an import statement here.
          """
          
          __all__ = ['LinearPackedParams', 'Linear']
          
          from torch.ao.nn.quantized.modules.linear import Linear
          from torch.ao.nn.quantized.modules.linear import LinearPackedParams