antoinedemathelin

Last batch should be of same length than others. We should use `.repeat` and then `.take` to get same length dataset with same size of batch for all batches. Additionaly,...

Some estimator based methods as FA can easily handle sparse array. Maybe KMM, KLIEP, LDM, NNW if pairwise distances can be computed

LDM, KMM tests fail on windows-latest

The use of ``tf.reduce_mean`` in train losses could lead to strange behavior. We should try to keep loss by sample and let tensorflow make the aggregation.

There are some bugs with mSDA

It may be a problem when using batchnorm for deep DA methods as the source and target data are given in separated batch.