LetsGoFir

Did you run the model on 8 GPUs or more? I cannot get a better result with more GPU, in fact it is worse.

 Why the PQ is not equal to SQ * RQ? For example, 0.78*0.524 = 0.40872 < 42.5.

'ggnn_epoch5_step14000.pt' have a a bit lower mIOU - 69

4

I change the data path, edit the path in ggnn.json, and test it with torch1.2/ptyhon3, got a mIoU 69, which is a bit lower than the result published in the...

About flickr30k training image number

1

Great job! I have a small question here, that you said below in your paper > Flickr30K consists of 31783 images collected from the Flickr website. Each image is accompanied...

Could you introduce more about model training?

3

Thanks for your excellent work! Hope for more details on training schedule, like training time, gpu number, and so on.

RuntimeError: The size of tensor a (32000) must match the size of tensor b (32001) at non-singleton dimension 0

8

fastchat==0.2.9 transformers is also the newest (commit a2789addd) vicuna-13b-delta-v1.1 

Will the performance be worse?

Any plan to upgrade transformers?

3

transformers=4.31.0 is a bit old

Finetuning on own data

1

Good Job! Do you have a plan to support LoRA or other PEFT?

chore: Only read cfg json once

1

Hello, I noticed that the code here repeatedly reads the JSON file, which doesn’t seem very efficient. Additionally, in some cases, this file may be deleted because the path is...