search

YBZh / Bridging_UDA_SSL

The official codes of "Semi-supervised Models are Strong Unsupervised Domain Adaptation Learners".
MIT License
35 stars 3 forks source link

test dataset #2

Open mompiou opened 4 days ago

mompiou commented 4 days ago

Thank you for this nice work. I was trying to apply SSL (Mean Teacher) for DA with my own dataset. I have synthetic data labelled (source), unlabelled data (target) and some labeled data from the target. Am I correct to use part of source target for validation set for weights update ? How to use labelled data from target ? What is the trans_test dataset in your code ?

YBZh commented 3 days ago

Am I correct to use part of source target for validation set for weights update ? --> I am not certain since I do not fully understand your question.

How to use labelled data from target ? --> The current code base seems not support the use of labeled target data. You need to implement it by yourself.

What is the trans_test dataset in your code ? --> In the transductive transfer learning setting, the unlabeled test data are used in the training process. Therefore, the trans_test is exactly the actual test data. In the inductive transfer learning setting, the trans_test and the actual test data are two different sampling from the same target domain.

mompiou commented 3 days ago

My main concern is how to deal with target data that are actually already labelled. As target data are unlabelled here I don't know how to use them (contrary to SSL where I use them in the supervised branch).

YBZh commented 3 days ago

Ok, I see. The current code base does not support three streams of input: labeled source data, labeled source data, and unlabeled source data. My suggestion is to add another "dataset" and corresponding "dataloader" in the prepare_data