Open seva100 opened 2 years ago
@seva100 have you had any luck?
@alaaobeid I used to email the authors, unfortunately, they don't have a checkpoint saved, so I had to retrain the model by myself. I wasn't able to reproduce the quality reported in the paper but tried to match it as much as possible. What I remember is that I had to fix the code to some extent and was unable to get a lot of improvement from using the attention mechanism and the adaptive layer, but it could be that they had to be applied somehow differently. Also, training was taking a lot of time in my case, but the authors told me that roughly 40 epochs should be enough to reach a sufficient quality level in practice.
@seva100 thanks for your reply. Would you be willing to share the fixed code and training setup? Any pointers would be greatly appreciated. Thanks in advance :)
Dear authors, thank you very much for releasing the repository. May I ask if a trained checkpoint for the ResNet-50 backbone can be provided? We would like to compare with your paper, but have been unable to run the code yet unfortunately. Thank you!