Hello author, I would like to ask you about the problem of model training code

[xiaodongqq]

Hello author, I am glad that you can reply to my previous questions. When I generated the corresponding key points and descriptors according to the readme you provided, I reported the following errors during training under Windows,

which could not be solved. Could you please tell me how to solve this problem? Thank you very much

[feixue94]

Hi, it seems like the input is None. can you print the values of input, e.g., keypoints, scores, and descriptors

[Uchan1996]

I got this error too. This error happens at the following code, which corresponds to the line 214 of the file layers.py. message = torch.einsum('bhnm,bdhm->bdhn', prob, value) The cause of this error is that the prob is None. Could you please fix the code to address this error?

[Uchan1996]

This error happens at the layer SharedAttentionalPropagation. So, If you train GM instead of DGNNS or AdaGMN, the training works.

[Uchan1996]

I got the following error, even if I train a GM. FileNotFoundError: [Errno 2] No such file or directory: '/data/datasets/object_recognition/sain/MegaDepth_undistorted/training_data/matches_sep_spp/0183/16115.npy' I think dump_megadepth.py does not generate 16115.npy.

[feixue94]

sorry for the late reply. I am a little bit busy these days and will test the code and fix bugs as soon as possible. Thank you very much your patience.

[feixue94]

I just fixed the bugs in training DGNNS and AdaGMN.

[feixue94]

I got the following error, even if I train a GM. FileNotFoundError: [Errno 2] No such file or directory: '/data/datasets/object_recognition/sain/MegaDepth_undistorted/training_data/matches_sep_spp/0183/16115.npy' I think dump_megadepth.py does not generate 16115.npy.

you might need to dump the training data first.

[Uchan1996]

I dumped the training data, which took a few days, and then I ran the train.py.

The assets/mega_scene_nmatches_spp.npy has the number of valid pairs for each scene. It says the scene 0183 has 16515 pairs, and the build_dataset_from_offline method in megadepth.py randomly samples 80 pair ids. However, the dump_megadepth.py generates only 15687 pairs in the folder of matches_sep_spp/0183. I think this is the cause of this error.

Could you please check whether the dumping and training code work correctly from the begining to the end, if you have chance.

[feixue94]

Dear all,

Sorry for the late reply. The code for dumping data and training model is tested and bugs are fixed now. The readme file is also updated. Please do a tiny test on 3 scenes for dumping data and training model before using full scenes for training because this could save your time if there is something wrong.

[feixue94]

I am going to close this. If you have any questions, feel free to reopen it.