search

MR-HosseinzadehTaher / BenchmarkTransferLearning

Official PyTorch Implementation and Pre-trained Models for Benchmarking Transfer Learning for Medical Image Analysis
Other
50 stars 14 forks source link

questions about parameters. #3

Open lichen14 opened 2 years ago

lichen14 commented 2 years ago

Hello, congratulations on the excellent paper! Can you share the parameters used for training in the args? I have reimplemented your work for DRIVE dataset, but I can not get your performance with ImageNet initial for pre-training. My result is mean dice=0.6985 with default args.

lichen14 commented 2 years ago

Sorry, I think I made some mistakes in training and above issue is solved. Another problem is that I got surprising performance after debugging (mean Dice = 79.69% for DRIVE segmentation) with the defaul args. (imagenet init) Are there any hidden errors in my trial?

MR-HosseinzadehTaher commented 2 years ago

For Drive, we used a batch size of 16. The reported performance for ImageNet is averaged over 10 runs.