Closed dahaigui closed 5 months ago
It seems to be a problem with your network, you can try using a proxy, or you can try downloading the weight locally, specify the local path and then running app.py.
Where should I put it in my local path?
I found the path I needed, $HOME/.cache/carvekit/checkpoints
Hi @dahaigui, is this resolved?
Hi @dahaigui, is this resolved?
Yes, this problem is solved.
Another question arise here
Is it possible to train with our own dataset?
As the "Could not create share link" message indicates, the Gradio share link error is due to your network connection. You can always use the local URL too.
Yes, to train on your own domain specific dataset you can simply replicate the format of our dataset, change the config path here, then execute the provided training command. You'd need occlusion images (256x256x3), corresponding whole (amodal) images (256x256x3), and binary visible (modal) masks (256x256x1).
The loader available here should further clarify the expected format. That said, instead of training from the Stable Diffusion weights (--finetune_from ckpt/sd-image-conditioned-v2.ckpt
), fine-tuning from our weights (--finetune_from ckpt/epoch=000005.ckpt
) can better preserve the generalization. This would highly depend on the # of samples in your domain specific dataset.
My local URL is not accessible, I've added port number 7860.
When I run the command
python app.py
, it gives me the error