SuperSupermoon / MedViLL

MedViLL official code. (Published IEEE JBHI 2021)
MIT License
89 stars 11 forks source link

About retrieval finetune question? #16

Open Subury opened 1 year ago

Subury commented 1 year ago

Can you share with the parameters setting during retrieval finetune? I try to use the define settings of code , but the results have very difference between the paper.

SuperSupermoon commented 1 year ago

@Subury Thanks for your interest. Could you please give me more detailed settings? As we described our paper and code, if you followed our code, the results should be the same. Therefore, I wonder your detailed settings. Then, we can catch the gap between us.

Subury commented 1 year ago

I think that the code on the "/MedViLL/downstream_task/retrieval/retrieval.py", which only has "Bidirectional" attention part. So, I use the "Bidrectional" pretrained weight to init finetune network, use the cmd of

"python retrieval.py --batch_size=18 --weight_load=True --load_pretrained_model=/MedViLL/pretrained/bi --epochs=50 --lr=1.8e-5"

to train.

In the inference phase, use the cmd of

"python retrieval.py --t2i=True --eval_len_size=1536 --do_train=False --do_test=True --label_conditioned_test_dataset=/dataset/MedViLL/T2I_Label_Test.jsonl --weight_load=True --load_pretrained_model=./output/2022-12-16\ 16\:16\:15.057888/49 --batch_size=160 --num_workers=8".

Thank you for your prompt reply.