Source code for talib.finetune.bss
"""
@author: Yifei Ji
@contact: jiyf990330@163.com
"""
import torch
import torch.nn as nn
__all__ = ['BatchSpectralShrinkage']
[docs]class BatchSpectralShrinkage(nn.Module):
r"""
The regularization term in `Catastrophic Forgetting Meets Negative Transfer:
Batch Spectral Shrinkage for Safe Transfer Learning (NIPS 2019) <https://proceedings.neurips.cc/paper/2019/file/c6bff625bdb0393992c9d4db0c6bbe45-Paper.pdf>`_.
The BSS regularization of feature matrix :math:`F` can be described as:
.. math::
L_{bss}(F) = \sum_{i=1}^{k} \sigma_{-i}^2 ,
where :math:`k` is the number of singular values to be penalized, :math:`\sigma_{-i}` is the :math:`i`-th smallest singular value of feature matrix :math:`F`.
All the singular values of feature matrix :math:`F` are computed by `SVD`:
.. math::
F = U\Sigma V^T,
where the main diagonal elements of the singular value matrix :math:`\Sigma` is :math:`[\sigma_1, \sigma_2, ..., \sigma_b]`.
Args:
k (int): The number of singular values to be penalized. Default: 1
Shape:
- Input: :math:`(b, |\mathcal{f}|)` where :math:`b` is the batch size and :math:`|\mathcal{f}|` is feature dimension.
- Output: scalar.
"""
def __init__(self, k=1):
super(BatchSpectralShrinkage, self).__init__()
self.k = k
def forward(self, feature):
result = 0
u, s, v = torch.svd(feature.t())
num = s.size(0)
for i in range(self.k):
result += torch.pow(s[num-1-i], 2)
return result
