Missing logger folder: result/task_pretrain_m3ae-seed0-from_

[trx14]

Hi, Thanks for your excellent work and share this awesome repo. When I run your code, I found a warning "Missing logger folder: result/task_pretrainm3ae-seed0-from" I checked the /result folder and it was empty. Do you have any suggestions for it? Thanks!

[zhjohnchan]

Hi there,

Thanks for your attention!

You can ignore that. It just tells us it will create the folder and store the results there.

Best, Zhihong

[trx14]

Thanks for your quick reply. I run the experiments on 4 A100 GPUs and changed the GPU nums to 4 in the shell. However, the program is stuck at the following step for 20 mins (see below figure), and the process didn't use the GPUs. Do you have any suggestions? Thanks!

[zhjohnchan]

Hi, here are some suggestions: (1) you can use 1 GPU for debugging so you can insert traces in your program; (2) then you can find the place it stuck. From my experience, it may be loading the data, so you can check if it is stuck by the memory limit.

[trx14]

Thanks for your suggestions, I run the experiments for VL-pertaining and fine-tune on VQA-SLAKE and IRTR-ROCO. Because of the memory limit, I only use ROCO for VL-pretraining. For SLAKE, I got a result of Open: 77% and Closed 85%. For IRTR-ROCO, I change the fine-tuning batch size to 20, I got a result for ir_r1: 6.5%, ir_r5: 21.7%, ir_r10: 33.5%, tr_r1: 7.0%, tr_r5:23.2% tr_r10: 34.9%. The IRTR performance is significantly lower than the paper. One explanation could be I only use ROCO for pretraining and the fine-tuning batch size is significantly smaller than your setting (which is 4096). I think your pretraining model can be a great baseline for many medical domain downstream tasks, and I really want to build some models based on your pretraining model, do you have any plan to open-source them? Thanks!

[zhjohnchan]

Hi, I think ROCO is too small to perform pre-training and try MedICaT together. For the pre-trained models, find them in the README, where I have already open-sourced.