Question about the inconsistent length of train_dataset and sample_weights #4
Description
Hello,
When I tried to reproduce the results for "domainnet real-clipart" using your code, I found an inconsistency in methods/ours/mi_dis_pm.py: len(train_dataset) and len(sample_weights) are not equal — they are 189818 and 187226, respectively.
I realized that this inconsistency may reflect the inconsistency of the subsampling result:
subsample_indices2 = subsample_balanced_instances(deepcopy(train_dataset2), prop_indices_to_subsample=0.3)
I carefully checked my DomainNet dataset against the one (Original version) provided at https://ai.bu.edu/M3SDA/, but I did not find any differences.
The number of images in my DomainNet dataset is:
| Clipart | Infograph | Painting | Quickdraw | Real | Sketch | |
|---|---|---|---|---|---|---|
| Num. | 48,834 | 53,201 | 75,759 | 172,500 | 175,327 | 70,386 |
Could you kindly provide your subsample_indices2, so that I can properly reproduce your results?
Here are the results on several other datasets：
| len(train_dataset) | len(sample_weights) | |
|---|---|---|
| cubc | 266724 | 266724 |
| fgvcc | 156652 | 156652 |
| scarsc | 369989 | 369989 |
| domainnet real-clipart | 189818 | 187226 |
| domainnet real-sketch | 196284 | 193524 |
Thank you in advance!