RL4M/PCRLv2

There are the following problems with fine-tuning brain data

liaochuanlin opened this issue · 4 comments

liaochuanlin commented

Namespace(data='/home/lcl/PCRLv2/dataset/BraTS_2018/MICCAI_BraTS_2018_Data_Training', model='pcrlv2', phase='finetune', b=4, epochs=100, lr=0.0001, output='./brats_finetune_weight', n='brats', d=3, workers=4, gpus='0,1', ratio=1.0, momentum=0.9, weight_decay='./pretrained_weight/simance_multi_crop_luna_pretask_1.0_240.pt', seed=42, amp=False)
pcrlv2_brats_finetune
Traceback (most recent call last):
File "/home/lcl/PCRLv2/main.py", line 46, in
data_loader = get_dataloader(args)
File "/home/lcl/PCRLv2/main.py", line 17, in get_dataloader
dataloader = getattr(generator, loader_name)()
AttributeError: 'DataGenerator' object has no attribute 'pcrlv2_brats_finetune'

Process finished with exit code 1

Luchixiang commented

For brats fine-tuning, to achieve better results, we modified based on a good baseline. I will send you the fine-tuning code through email. Please try fine-tuning based on that code.
Maybe you should first adjust the hyper-parameters to achieve comparable training from scratch results and then try our pretrained weight.

funnyzhou commented

Hi @liaochuanlin,

Doing segmentation on BraTS is not easy considering there are so many details to deal with.

Please try matching the performance of the training from scratch baseline first.

liaochuanlin commented
liaochuanlin commented