surrey-up-lab / regionspot Goto Github PK

在使用自己的数据集进行训练时，我发现需要一个offline_token：
#read pth
pth_file = os.path.join(self.mask_tokens_dir, os.path.join(dataset_name, str(image_id)+'.pth'))
offline_token = torch.load(pth_file)
请问是需要提前在SAM中得到每个物体的分割embeddings吗？

_PREDEFINED_SPLITS_OPENIMAGES = {
"openimages_train": ("openimages/detection/", "re_openimages_v6_train_bbox_splitdir_int_ids.json"),
"openimages_val": ("openimages/detection/", "re_openimages_v6_train_bbox_splitdir_int_ids.json"),
}

When I performed the second phase of training, loading v3det_train was stuck. After pausing the program, an error like this appeared: AssertionError: Attribute 'thing_classes' in the metadata of 'v3det_train' cannot be set to a different value! This is why?

Thanks for the great work!

Have you done any experiment on different sam model such as same_vit_h_4b8939.pth? What would the result be? Any improvement?

Thanks
George

It seems that the checkpoints are placed at the netdisk of the University of Surrey, where I could not ascess

您好，请问你们在实验中，SAM的prompt输入是point还是bbox呀

Amazing work! When could the training code be open-sourced? I'm eager to follow your work! Thanks!

I find point and bbox as prompt but no text, is it possible to support text as prompt？

Hello, what's the prompt of SAM when in inference? And what fps can this model achieve? Thanks

the model download link you provided is not granted to access, or resource is moving?

你好，请问你们最近会公布训练脚本吗
我在研究train_net的代码时，好像没有发现从RegionSpot()方法中调用模型的部分

OpenImages has 561GB. Do I need to download a subset of it or all of it?

Amazing work! How to dowlodad the checkpoint??

Excuse me about how long the training cycle is?