The GPUs for training Text/Visual Embedder

[Cyyyyyb]

Hi @gy65896

Thanks for your nice work.

I am trying to reproduce your method from scratch.

I would like to know what GPU the text/visual embedder is trained on.

Best regards and many thanks,

[gy65896]

Thanks for your attention!

Due to the small GPU memory consumption compared with the restoration model, we just used a single 3090 GPU to train text and visual embedded.

[Cyyyyyb]

Thanks for your attention!

Due to the small GPU memory consumption compared with the restoration model, we just used a single 3090 GPU to train text and visual embedded.

Got it, thanks for the quick response!

However, when I used a single 3090 GPU to train text and visual embedded, I reported an error and would like to ask you for help.

Best regards and many thanks,

I use this line command to run the code: python train_Embedder.py --train-dir ./data/CDD-11_train --test-dir ./data/CDD-11_test --check-dir ./ckpts --batch 256 --num-workers 0 --epoch 200 --lr 1e-4 --lr-decay 50

[gy65896]

Have you attempted to set batch_size as 128? Perhaps I use this set of parameters on another device.

[Zhong-Polyu]

Thanks for your attention! Due to the small GPU memory consumption compared with the restoration model, we just used a single 3090 GPU to train text and visual embedded.

Got it, thanks for the quick response!

However, when I used a single 3090 GPU to train text and visual embedded, I reported an error and would like to ask you for help.

Best regards and many thanks,

I use this line command to run the code: python train_Embedder.py --train-dir ./data/CDD-11_train --test-dir ./data/CDD-11_test --check-dir ./ckpts --batch 256 --num-workers 0 --epoch 200 --lr 1e-4 --lr-decay 50

Hi, I encountered the same problem. I found it caused by the memory cost at the validation stage. Add the ''with torch.no_grad():'' after model.val(), solved!

[gy65896]

Thank you for your mention. I will update the code.