The detail of ProDe

[a12541115336]

First Congrats for accepted by ICLR！Recently, I replicated ProDe on Domainnet126. However, results are far from what the paper suggests (about 5%-10%). Besides, I find that the backbone claimed in the paper which is different in the yaml file. resnet50 in yaml but resnet101 in paper. Could you provide with the checkpoints for ProDe on Domainnet126 dataset?

[Kalaniik]

Thank you very much for raising this issue and for the careful effort you put into reproducing our results.

We address your concerns point by point:

1. Reproducibility of ProDe on DomainNet-126
During the process of integrating ProDe into our unified framework for prior methods, an implementation oversight occurred: a subset of network modules was initialized with unintended weights, i.e., the weight initialization for part of the network deviated from the intended specification. We have now identified and corrected the relevant code paths, and the repository has been updated accordingly.
To make the above modification easy to verify and to support precise configuration alignment, we have also uploaded the corresponding training logs for ProDe on DomainNet-126（See domainnet126.zip below）. These logs provide an auditable reference of the end-to-end protocol, helping pinpoint any remaining sources of variance.

2. Backbone inconsistency (YAML vs. paper)
You are absolutely correct that there is an inconsistency between the backbone specified in the YAML configuration and that stated in the manuscript. This is an oversight during the paper-writing process. We will revise the paper accordingly to ensure the backbone specification is consistent and unambiguous. Thank you for catching this important detail.

We appreciate your careful verification and constructive feedback—issues like this are crucial for improving clarity and ensuring reproducibility.

domainnet126.zip