Closed Jyerim closed 1 year ago
The code has been updated by @GioFic95, so the batch_size and learning rate should also be updated to achieve a better result.
The results reported in issue #26 are obtained by using the parameters described in the paper. I.e., I ran the finetuning with this command:
FSC_finetune_cross.py --epochs 1000 --batch_size 8 --lr 1e-5 --output_dir ./data/out/finetune --log_dir None --title CounTR_finetuning_paper --resume ./data/out/pretrain/checkpoint__pretraining_299.pth
Moreover, in my changes I haven't touched the default values for batch size, learning rate or gradient accumulation, and neither the training itself. I just added another logger alongside tensorboard: W&B. Thus, the hyperparameters shouldn't need any updates.
Sorry for my inconsistencies of the hyperparameters.
I modified the finetuning code to conduct ablation studies and forgot to recover it completely. So for all hyperparameters except the learning rate, please refer to the original paper. For the learning rate, I used batch learning rate in the code and calculated the corresponding learning rate myself in the paper. So the batch learning rate in the code is correct.
For learning rate, I use a batch learning rate of 2e-4 as I mentioned in the readme.md. For batch size, I use 8, though I think a larger batch size will yield a better result.
Hope my description will help you.
Thank you for your answer.
Are batch_size and learning rate the same as stated in the paper during Fine-tuning stage? In the script file, batch 26, lr=2e-4 is set to be learned. On the other hand, the paper states batch 8, lr=1e-5. This is the result of finetune with the provided FSC147.pth. Batch 26, lr=2e-4 : 12.79 mae/ 86.49 rmse Batch 8, lr=1e-5 : 13.77 mae/ 87.69 rmse