Yufei Xu
Yufei Xu
Hi, ViTPose has a smaller training memory than HRNet-w48 when inputting 256x192 images. Perhaps you could try resizing the input images for your own data to alleviate the training memory...
This is the default post-processing hyper-parameters to reduce duplicate predictions.
It seems there are no further questions. I will close the issue temporarily. If you have any further questions, please feel free to re-open it.
Hi, There is an initial version of the training code [link](https://drive.google.com/file/d/1__KXQRfKHsyJdXAjh0g4oO3BUxpCmgT5/view?usp=share_link). It is temporally not suitable to be made public here since it lacks a readme file and proper comments...
> The issue It seems that there is no mask data in the dataset directory. If the mask data is not available or expected.to use, please disable the mask loading...
Maybe you forgot to disable the mask prediction in the model definition?
Hi, you can change the default master port by importing `PORT=XXX` or explicitly setting `--master_port=xxxx` to find an unused port.
Hi, Thanks for your interest and we have not tried weights using other SSL methods yet.
The pretrained models should be loaded by specifying the `model.pretrained` options, not the `resume_from` options. Please refer to [Usage](https://github.com/ViTAE-Transformer/ViTDet#usage) for detailed commands.
Hi, Please refer to Sec 3.3 and Sec 4.4.1 in the [ViTAEv2](https://arxiv.org/pdf/2202.10108.pdf) paper. We use MAE to train the isotropic design of ViTAE, not the hierarchical ViTAEv2 backbone. We will...