I found if a dataset contains cells less than 1/4 the number of cells in the other dataset, the algorithm would train without using the smaller dataset at all. Look at line 142 for batch_size assignment in scDART.py and line 151 in train.py.

Also I am curious about the way you set bacth_size (1/4 of the larger dataset), would that be unnecessarily too large for dataset contain hundreds of thousands of cells in terms of memory usage? Will that be a concern if we fix batch_size at smaller number like 128 (this is what I saw other auto-encoder based algorithms used)?