xref: /aosp_15_r20/external/pytorch/torch/ao/pruning/_experimental/pruner/parametrization.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1# mypy: allow-untyped-defs
2import torch
3from torch import nn
4from torch.nn.utils.parametrize import is_parametrized
5
6
7def module_contains_param(module, parametrization):
8    if is_parametrized(module):
9        # see if any of the module tensors have a parametriztion attached that matches the one passed in
10        return any(
11            any(isinstance(param, parametrization) for param in param_list)
12            for key, param_list in module.parametrizations.items()
13        )
14    return False
15
16
17# Structured Pruning Parameterizations
18class FakeStructuredSparsity(nn.Module):
19    r"""
20    Parametrization for Structured Pruning. Like FakeSparsity, this should be attached to
21    the  'weight' or any other parameter that requires a mask.
22
23    Instead of an element-wise bool mask, this parameterization uses a row-wise bool mask.
24    """
25
26    def __init__(self, mask):
27        super().__init__()
28        self.register_buffer("mask", mask)
29
30    def forward(self, x):
31        assert isinstance(self.mask, torch.Tensor)
32        assert self.mask.shape[0] == x.shape[0]
33        shape = [1] * len(x.shape)
34        shape[0] = -1
35        return self.mask.reshape(shape) * x
36
37    def state_dict(self, *args, **kwargs):
38        # avoid double saving masks
39        return {}
40
41
42class BiasHook:
43    def __init__(self, parametrization, prune_bias):
44        self.param = parametrization
45        self.prune_bias = prune_bias
46
47    def __call__(self, module, input, output):
48        if getattr(module, "_bias", None) is not None:
49            bias = module._bias.data
50            if self.prune_bias:
51                bias[~self.param.mask] = 0
52
53            # reshape bias to broadcast over output dimensions
54            idx = [1] * len(output.shape)
55            idx[1] = -1
56            bias = bias.reshape(idx)
57
58            output += bias
59        return output
60