First, I want to say thank you for your awesome product!
I want to utilize your code in uncond mode for the specific purpose of human skeleton reconstruction. However, I encountered an issue while using the code. I followed what you uploaded on Git and trained as follows:
I have a guess as to what this might be. Can you try and generate motions with full length (196) instead of the default 120 and let me know what happens?
Hello!
First, I want to say thank you for your awesome product!
I want to utilize your code in uncond mode for the specific purpose of human skeleton reconstruction. However, I encountered an issue while using the code. I followed what you uploaded on Git and trained as follows:
python -m train.train_mdm --save_dir save/my_name5 --dataset humanact12 --cond_mask_prob 0 --lambda_rcxyz 1 --lambda_vel 1 --lambda_fc 1 --unconstrained --train_platform_type TensorboardPlatform
And for inference, I ran:
python -m sample.generate --model_path ./save/my_name5/model000300000.pt --num_samples 4 --num_repetitions 3
However, the generated motion appears to be static, as if it's producing only one pose. I am attaching a video for better understanding.
Did I make a mistake in executing your code? Or are there additional considerations when using an unconditional model?
Thank you.
https://github.com/GuyTevet/motion-diffusion-model/assets/47049325/aaa48b8c-d63c-46b3-a9da-b3dc4ea5ae77