tencentailabhealthcare / concl Goto Github PK

Hi, thanks for sharing this work. I wonder during the finetuning stage, do you use any annotations from Glas/CRAG dataset, or just finetune the detection head in SSL manner?

Thanks for sharing your great work!
In the paper, you mentioned that TCGA-BRCA-100K is used for pre-training, whereas it seems not covered in your public code?

Hi @Jiawei-Yang,

Thanks for the great work and sharing your code!

I tried to reproduce your training results. I downloaded your training dataset and followed your instructions to train on 2 GPUs using the command "bash tools/dist_train.sh configs/NCT_pretraining/concl/r18_bs256_ep200.py 2" (I set batch_size to 64 and imgs_per_gpu= batch_size//2).
The strange thing is that the training loss has not been reduced, and it only takes 10 hours to train 200 epochs.
The initial loss in epoch 1: loss: 7.0236, instance_l: 7.0236, concept_l: 0.0000
The loss in the end of epoch 200: loss: 8.2989, instance_l: 9.6324, concept_l: 8.2989
Attached is my training log.

Could you guess what went wrong?
train.log