This is the PyTorch implementation of the Weighted Batch Normalization layers used in Boosting Domain Adaptation by Discovering Latent Domains. The layer is composed by a native CUDA implementation plus a PyTorch interface.

This version has been tested on:

• PyTorch 0.4.1
• python 3.6
• cuda 9.2

Note: This code have been produced by taking large inspiration from the In-Place Activated BatchNorm implementation.

The only requirement for the layer is the cffi package. It can be easily installed by just typing:

pip install cffi

Before using the layer, it is necessary to compile the CUDA part. To compile it, just type:

sh build.sh
python build.py

To check that the installation runs fine, just type:

 python test_wbn.py

Three layers are available:

• WBN2d : the 2D weighted BN counterpart of BatchNorm2d. It can be initialized specifying an additional parameter k denoting the number of latent domains (default 2). As input during the forward pass, for each sample in the batch it takes an additional tensor w denoting the probability that the sample belongs to each of the latent domains.
• WBN1d : the 1D weighted BN counterpart of BatchNorm1d. The initialization and forward pass follows WBN2d.
• WBN : the single domain weighted batch norm. This layer is the base component for the previous two. It does not take k as additional initialization input. Moreover, the tensor w given as additional input has a number of element equal to the batch-size and must sum to 1 in the batch dimension.

The layer is initialized as standard BatchNorm except that requires an additional parameter regarding the number of latent domains. As an example, considering an input x of dimension NxCxHxW where N is the batch-size, C the number of channels and H, W the spatial dimensions. The WBN2d counterpart of a BatchNorm2d would be initialized through:

wbn = wbn_layers.WBN2d(C, k=D) 

with D representing the desired number of latent domains. Differently from a standard BatchNorm, this layer will receive an additional input, a vector w of shape NxD:

out = wbn(x, w) 

w represents the probability that each sample belongs to each of the latent domains thus it must sum to 1 in the domain dimensions (i.e. dim=1). Notice that for the WBN case w has dimension Nx1 and must sum to 1 in the batch dimension (i.e. dim=0).

If you find this code useful, please cite:

@inProceedings{mancini2018boosting,
author = {Mancini, Massimilano and Porzi, Lorenzo and Rota Bul\`o, Samuel and Caputo, Barbara and Ricci, Elisa},
title  = {Boosting Domain Adaptation by Discovering Latent Domains},
booktitle = {Computer Vision and Pattern Recognition (CVPR)},
year      = {2018},
month     = {June}
}