Problem about the implementation of lmpt/csel.py.

[richard-peng-xia]

Problem about the implementation of lmpt/csel.py.

[richard-peng-xia]

Hi @Wang-zuozheng,

Thanks for your interest in LMPT!

1. Thank you for reminding me. In fact, it should be written as class_weights = sfm(torch.from_numpy(np.asarray(mmcv.load(freq_file)['class_freq'])).to(torch.float32).cuda()), where I forgot to update the code. Details can be found here.

2. It is not an addition softmax operation. This line of code corresponds to Eq. 5, which is the re-weight operation.

3. This error should not happen as there are corresponding images for each category, although the number varies.

I hope it may help you. Please feel free to ask if there is any further problem.

Thank you very much!