Open JungHunOh opened 5 months ago
Thanks for flagging this issue, which we take seriously. The first authors Haoyue and Yifei will follow up.
Thanks for your answer. Please be aware that the training log is currently not shown since I deleted them by mistake in the wandb server. Instead, I share the log file directly here. output.log
Hello, Thank you for your interest in our work. The script/train_hypo_dg.sh file includes default hyperparameters. However, we conduct hyperparameter tuning following common practice in DomainBed. The optimal hyperparameters vary by domain; for example, the best 'lr' for the cartoon domain is 0.0005, with 'batch_size' of 32 and 'w' of 4.0. Please refer to the appendix of our paper for a detailed range of hyperparameters. We also updated our script for the hyperparameters of each domain.
Hello,
I'm currently working on replicating the results from the PACS dataset using your code.
However, I've encountered a discrepancy as the obtained accuracies fall significantly below those reported in the paper.
For instance, when targeting the 'cartoon' domain, I achieved an accuracy of 79.78% on the PACS dataset. (reported accuracy = 82.3%)
I executed the script/train_hypo_dg.sh file.
I've not checked if there are similar issues on the other datasets.
Here is the training log: https://wandb.ai/junghunoh/hypo/reports/PACS-target-domain-cartoon---Vmlldzo3NDY5NTkw
Here are the training arguments. {'augment': True, 'batch_size': 64, 'bottleneck': True, 'cosine': True, 'epochs': 50, 'feat_dim': 512, 'gpu': 0, 'head': 'mlp', 'id_loc': 'datasets/PACS', 'in_dataset': 'PACS', 'learning_rate': 0.0005, 'loss': 'hypo', 'lr_decay_epochs': '100,150,180', 'lr_decay_rate': 0.1, 'mode': 'online', 'model': 'resnet50', 'momentum': 0.9, 'normalize': False, 'prefetch': 4, 'print_freq': 10, 'proto_m': 0.95, 'save_epoch': 100, 'seed': 4, 'start_epoch': 0, 'target_domain': 'cartoon', 'temp': 0.1, 'trial': '0', 'use_domain': False, 'w': 2.0, 'warm': False, 'weight_decay': 0.0001}
I would appreciate if you check this issue.
Thanks.