There might be a minor bug in data.StructureLoader. Some samples will not be included in the loader, so the dataset size is not exactly the same as that reported in the paper.
for ix in sorted_ix:
size = self.lengths[ix]
if size * (len(batch) + 1) <= self.batch_size:
batch.append(ix)
batch_max = size
else:
clusters.append(batch)
batch, batch_max = [], 0
There might be a minor bug in data.StructureLoader. Some samples will not be included in the loader, so the dataset size is not exactly the same as that reported in the paper.