fhaghighi / dira Goto Github PK

This is a great job, which unifies contrastive and generative method. And writting is also good. But my intuition tells me that this method seems sensitive to hyperparameters.

I wonder what is the purpose or function of the class 'Custom_Distoration' in the transformation.py file?

hi, thansk for open-source.
can you share the pre-trained weights? such as resnet50 backbone and others.
thanks.

I think it is a great job for medical image analysis! However, I have some questions about the details of training 2D images.
①You did a warming up and then jointly training, which totally took up to 800 epoches as shown in your paper. Could you please tell me how many epoches it takes for warming up and how many epoches for jointly training?
②In the jointly training stage, you set the batch size to 16 instead of 256 in the repository, but it seems not to be mentioned in your paper. But I found it take little gpu memory when I set batch size to 16 and work well for batch size of 256. Is there any reason to set the batch size smaller? What happened if larger batch size like 256 is used during jointly training?
Thanks in advance for your time!

Thank you very much for sharing the codes. Could you please share the code for training the 3D models by applying DIRA to TransVW? If already available, could you please provide the link to the shared code?