Pretrained Pose-to-body Model? about vid2vid

[tcwang0509]

I'm not sure we can release that model due to copyright issue.
The training code is committed now, so you're welcome to try it yourself.

[petergerten]

@tcwang0509 It would be great if the model could be released (or maybe a new model with training data that can be used).

As most people won't have a DGX1 available, training the 2K resolution net for 10 days on a p3.16xlarge instance (also has 8 V100 GPUs) would cost USD 5875 on AWS. (USD24.48 per hour on-demand pricing * 24 hours/day * 10 days)

And it seems that this won't even work as you state at least 24GB memory is required per GPU. I am not even aware of any cloud provider currently offering GPUs with more than 16GB.

[tcwang0509]

@petergerten the requirements you mentioned is for training Cityscapes. For pose even on single GPU it should just take 5-7 days.

[petergerten]

@tcwang0509 great, thanks for the clarification

[ChenyuGao]

The research is very exciting!
I also hope the pretrained model will be released~

[bube5h]

@tcwang0509 For face how many days will it take?

[therobotprogrammer]

So I've found a pre-trained model used by densepose.
https://dl.fbaipublicfiles.com/densepose/DensePose_ResNet101_FPN_s1x-e2e.pkl

Now, if only there was a way to convert this densepose model to a format used by vid2vid. Assuming both libraries use the same formatting to declare the input & output tensors, would this be possible?

I'm new to pytorch so please pardon the noob question.

[kartikJ-9]

So I've found a pre-trained model used by densepose.
https://dl.fbaipublicfiles.com/densepose/DensePose_ResNet101_FPN_s1x-e2e.pkl

Now, if only there was a way to convert this densepose model to a format used by vid2vid. Assuming both libraries use the same formatting to declare the input & output tensors, would this be possible?

I'm new to pytorch so please pardon the noob question.

Any update on this one. This seems a good approach to get the pretrained model for pose?