import torch
import torch.nn as nn
__all__ = ['MinibatchDiscrimination1d']
# The original paper by Salimans et. al. discusses only 1D minibatch discrimination
[docs]class MinibatchDiscrimination1d(nn.Module):
r"""1D Minibatch Discrimination Module as proposed in the paper `"Improved Techniques for
Training GANs by Salimans et. al." <https://arxiv.org/abs/1805.08318>`_
Allows the Discriminator to easily detect mode collapse by augmenting the activations to the succeeding
layer with side information that allows it to determine the 'closeness' of the minibatch examples
with each other
.. math :: M_i = T * f(x_{i})
.. math :: c_b(x_{i}, x_{j}) = \exp(-||M_{i, b} - M_{j, b}||_1) \in \mathbb{R}.
.. math :: o(x_{i})_b &= \sum_{j=1}^{n} c_b(x_{i},x_{j}) \in \mathbb{R} \\
.. math :: o(x_{i}) &= \Big[ o(x_{i})_1, o(x_{i})_2, \dots, o(x_{i})_B \Big] \in \mathbb{R}^B \\
.. math :: o(X) \in \mathbb{R}^{n \times B}
This is followed by concatenating :math:`o(x_{i})` and :math:`f(x_{i})`
where
- :math:`f(x_{i}) \in \mathbb{R}^A` : Activations from an intermediate layer
- :math:`f(x_{i}) \in \mathbb{R}^A` : Parameter Tensor for generating minibatch discrimination matrix
Args:
in_features (int): Features input corresponding to dimension :math: `A`
out_features (int): Number of output features that are to be concatenated corresponding to dimension :math: `B`
intermediate_features (int): Intermediate number of features corresponding to dimension :math: `C`
Returns:
A Tensor of size :math: `(N, in_features + out_features)` where :math: `N` is the batch size
"""
def __init__(self, in_features, out_features, intermediate_features=16):
super(MinibatchDiscrimination1d, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.intermediate_features = intermediate_features
self.T = nn.Parameter(torch.Tensor(in_features, out_features, intermediate_features))
nn.init.normal_(self.T)
[docs] def forward(self, x):
r"""Computes the output of the Minibatch Discrimination Layer
Args:
x (torch.Tensor): A Torch Tensor of dimensions :math: `(N, infeatures)`
Returns:
3D Torch Tensor of size :math: `(N,infeatures + outfeatures)` after applying Minibatch Discrimination
"""
M = torch.mm(x, self.T.view(self.in_features, -1))
M = M.view(-1, self.out_features, self.intermediate_features).unsqueeze(0)
M_t = M.permute(1, 0, 2, 3)
# Broadcasting reduces the matrix subtraction to the form desired in the paper
out = torch.sum(torch.exp(-(torch.abs(M - M_t).sum(3))), dim=0) - 1
return torch.cat([x, out], 1)