Linjie Li

One more piece of information. From our past experience, if the loss scaler stay above 1, the training should be steady.

Our docker is built on V100 GPUs and inherited from [nvidia built pytorch docker](https://github.com/linjieli222/HERO/blob/32c1c523c7a9f547a29f14c8e33dec24ebd14156/Dockerfile#L1): `pytorch:19.10-py3` from [here](https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel_19-10.html#rel_19-10). If you wish to support on your own GPU version, you may need...

If `pytorch:19.10-py3` is compatible with RTX 3090, you can simply run `docker build ./ -t ` to rebuild the same enviroments to the released docker image.

Thanks for your interests in this project. Please take a look at https://github.com/linjieli222/HERO#quick-start. When launching docker container, you will need to point to $PATH_TO_STORAGE

Take look at this file https://github.com/linjieli222/HERO/blob/32c1c523c7a9f547a29f14c8e33dec24ebd14156/launch_container.sh#L23 Basically, we map (1) $PATH_TO_STORAGE/txt_db to /txt inside the docker (2) $PATH_TO_STORAGE/video_db to /video inside the docker (3) $PATH_TO_STORAGE/finetune to /storage inside the docker...

The simple answer is 8 x the default value in config. However, it is highly possible that your single-GPU finetuning cannot match our 8-GPU finetuning result. Check this thread on...

Please refer to the docker file [here](https://github.com/linjieli222/HERO/blob/master/Dockerfile).

I just re-tested the connect and could not reproduce the error. I suspect it is due to unstable internet connection.  There is not much we can do from our...

@nikhilbchilwant @ifmaq1 @dami23 if you are still encountering issues with downloading. I suggest you to try download with [azcopy](https://docs.microsoft.com/en-us/azure/storage/common/storage-use-azcopy-v10). An example command to download the _pretrained_models.zip_ is: ``` cp "https://convaisharables.blob.core.windows.net/vqa-regat/pretrained_models.zip"...

Hi, thanks for your interests in this work and sorry for the late reply. They are one-hot embeddings.