omegahh
commented
3 years ago Thank you for point out the issue. Since the analyzing codes are abundant for the core codes for the model. I only extracted the core codes to be this stand-alone repository. But it was still changed many times which makes you confused. I am sorry about that.
View the commit history. In 26 June 2019, I committed a change which unintentionally changed the train-test split (maybe I wanted to test the training step works well in different settings). While you can see that the models in the '/save' file were changed in 15 Aug. 2019.
In this situation. I think the models may be trained on chr[1, 2, 3, 4, 5, 6, 7, 12, 13, 14, 15, 16, 17, 18]. I am not pretty sure, but the training code is published. The model base on chr1-14 can be reproduced. If you want the model directly from me. I am happy to retrain it for you. But I can not guarantee the exact time because I am not in the Beijing lab this year.
About training epochs, I recommend you that training DeepHiC more than 200 epochs. In my experience, 200-epoch-training is sufficient for the model, but training in 300 even 400 epochs is permitted which may get a bit of performance improvement.
avsarg
Hi,
I am trying to work with DeepHiC, but I am a bit confused.
In the publication and also in the summary part in github, it is indicated that the training and prediction are performed on chr 1-14 and chr 15-22, respectively. And, 300 epochs are used.
On the other hand, the scripts here use 200 epochs, and "training on chr [1, 2, 3, 4, 5, 6, 7, 12, 13, 14, 15, 16, 17, 18]" and "testing on chr [8, 9, 10, 11, 19, 20, 21, 22]" are performed.
I just want to know that which preferences are used for the training of the models that are located in the "/save" file?
Thanks!