how can i ues the trianed pth to the origin supergluepretrain code
Opened this issue · 4 comments
ZhouShunlong commented
how can i ues the trianed pth to the origin supergluepretrain code
Shuhul24 commented
Hey @ZhouShunlong did you get any way how to use the weights that we get by training the superglue with the pretrained model?
ZhouShunlong commented
It seems that we can't get any details about how superglue train it's model
…---Original---
From: ***@***.***>
Date: Tue, Nov 7, 2023 14:28 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [yingxin-jia/SuperGlue-pytorch] how can i ues the trianed pth to the origin supergluepretrain code (Issue #29)
Hey @ZhouShunlong did you get any way how to use the weights that we get by training the superglue with the pretrained model?
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>
Shuhul24 commented
Well what I observed is that there are some changes in the architecture that is being used in official SuperGlue such as in line 198 in models/superglue.py the default config are different than the one in the official one.
Also in line 249 (in models/superglue.py) the script is
desc0 = desc0 + self.kenc(kpts0, torch.transpose(data['scores0'], 0, 1))
desc1 = desc1 + self.kenc(kpts1, torch.transpose(data['scores1'], 0, 1))whereas in the official models/superglue.py it is
desc0 = desc0 + self.kenc(kpts0, data['scores0'])
desc1 = desc1 + self.kenc(kpts1, data['scores1'])I tried importing the trained model weights in the official model but it wasn't working as there were plenty of errors coming around these lines.
ZhouShunlong commented
Can you speak in Chinese ？We can talk in WeChat with searching my phone number ：18800543847
…---Original---
From: ***@***.***>
Date: Wed, Nov 8, 2023 13:35 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [yingxin-jia/SuperGlue-pytorch] how can i ues the trianed pth to the origin supergluepretrain code (Issue #29)
Well what I observed is that there are some changes in the architecture that is being used in official SuperGlue such as in line 198 in models/superglue.py the default config are different than the one in the official one.
Also in line 249 (in models/superglue.py) the script is
desc0 = desc0 + self.kenc(kpts0, torch.transpose(data['scores0'], 0, 1)) desc1 = desc1 + self.kenc(kpts1, torch.transpose(data['scores1'], 0, 1))
whereas in the official models/superglue.py it is
desc0 = desc0 + self.kenc(kpts0, data['scores0']) desc1 = desc1 + self.kenc(kpts1, data['scores1'])
I tried importing the trained model weights in the official model but it wasn't working as there were plenty of errors coming around these lines.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you were mentioned.Message ID: ***@***.***>