How do I use the pretrained Matting checkpoints for training?

[tarunn2799]

I've been training my own MODNet models on custom data, using the configs provided. How do I use the model checkpoint provided in the readme HERE as a starting point for my training?

Also, how should I change my training params if I'm using this as a checkpoint to start training? I currently use the default config for Modnet-hrnet18, and I get good results. Looking to make it better.

Final question : the HRNet backbone that is being used for modnet, is it pretrained on ImageNet or is it from scratch?

[wuyefeilin]

For the first question, you can add pretrained: path/to/xxxx.pdparams' of the model` in configs as following:

Final question: It is pretrained on ImageNet.

May I ask, in what business do you use matting？

[tarunn2799]

Hi @wuyefeilin I did try doing that - I changed the outer pretrained key under the model key to the downloaded MODNet HR18 .pdparams file. But I got ValueError: paddle.load can not parse the file: <the downloaded file path>

Turns out the file was corrupted, and upon retrying it from scratch, it worked. Thank you. Can you help me understand if I should change the lr, optimizer and num_iters now since I'm using a pretrained model?

One more question : Will you guys be supporting the SOC adaptation strategy in MODNet within PaddleSeg anytime soon? Any idea how I can try that out with the trained models?

Your final question : I don't exactly use it for any business, personal use for my own learning and projects.

[wuyefeilin]

I download the modnet-hrnet_w18.pdparams form https://paddleseg.bj.bcebos.com/matting/models/modnet-hrnet_w18.pdparams can run normally. May be you can try again.

You can reduce the lr and num_iters if you using a pretrained model

We have not supported SOC. If you are interesting, may be you can use it on you unlabeled dataset by fix the loss module.