Open Rui-Zhou-2 opened 7 months ago
Yes, you need to start by generating pseudo-labels and training an initial model using the target domain images and these pseudo-labels. Although the model at this stage may have some errors, its performance will gradually improve through later fine-tuning. Even though pseudo-labels are not entirely accurate, the model can still learn some useful information.
Hi, may I ask how many epoches you train A2B_retrain(using the target(B) domain images and pseudo-labels and A2B_fintune(using the selected) separately? I am getting pretty bad results on A2B_retrain IMG_85.jpg Gt 5.50 Pred 469 IMG_99.jpg Gt 10.00 Pred 504 [2024-02-13 23:26:46] [32mIntermediate result: 877.1661392405064 (Index 6)[0m
Hi, may I ask how many epoches you train A2B_retrain(using the target(B) domain images and pseudo-labels and A2B_fintune(using the selected) separately? I am getting pretty bad results on A2B_retrain IMG_85.jpg Gt 5.50 Pred 469 IMG_99.jpg Gt 10.00 Pred 504 [2024-02-13 23:26:46] �[32mIntermediate result: 877.1661392405064 (Index 6)�[0m
- MAE 877.166
- MSE 947.284
Please check your code, especially whether you are using the correct checkpoint file. There must be an issue somewhere, as we have never encountered this situation during our training process.
Hi, I am wondering if I want to retrain the model, would it be right to run pseudo_generate first? Since the outcome is mostly wrong And what is the setting for A2B_retrain_root and A2B_fineturn_root ?