descriptor loss

[Hjy0801]

Hello, when i trained the full pipeline, the descriptor loss can not decreased. it was about 1 all time. why is this? Can you help me? Thank you.

[rpautrat]

Hi, can you share the tensorboard curves of the different losses? Are the other losses decreasing or not?
On which dataset are you training also?

[Hjy0801]

Sorry, I don't save the tensorboard curves now. I just trained it for several hours and I use the sold2_synthetic.tar as the pre-train model. i train it on wire-frame datasets(exported) .The other losses can decrease. But the descriptor loss can not decrease.  The full train process spends too much time. so i don't train it fully. I want to to know whether it is right.

…

------------------ Original ------------------ From: ***@***.***>; Date:  Wed, Mar 15, 2023 09:25 AM To: ***@***.***>; Cc: ***@***.***>; ***@***.***>; Subject:  Re: [cvg/SOLD2] descriptor loss (Issue #78)   Hi, can you share the tensorboard curves of the different losses? Are the other losses decreasing or not? On which dataset are you training also? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[rpautrat]

Did you change anything in the code of the main branch?
Did you try visualizing the exported dataset with this notebook to check that everything was fine?

[Hjy0801]

I did't change any code and check the exported dataset. I operated the command of step 5 and used the sold2_synthetic.tar as the pre-train model.

This is the train process. descriptor_loss always is 1.

This is the train datasets.

[rpautrat]

The exported dataset looks good.

Did you modify the field 'return_type' in config/wireframe_dataset.yaml to 'paired_desc', as requested in the ReadMe?

[Hjy0801]

yes,I did. when i use the sold2_wireframe.tar as pre-train model. the descriptor_loss are very low and the descritor_loss changed normally.

[rpautrat]

In our original work, we first did step 4 of the Readme, before proceeding to step 5. But as I understand it, you skipped step 4 (i.e. training the detector only on the Wireframe dataset), right? Maybe that is why the descriptor loss is stuck, you need to pre-train the network with step 4, then fine-tune it with the descriptor in step 5.

The value of 1 displayed in your descriptor loss is basically the margin used in the triplet loss. In my experience, it is common that when training a network with triplet loss, the descriptor loss can stay stuck at this margin value (1 in your case) for a while. But if you keep training, the network eventually manage to improve and to go below this margin value.

So as a conclusion, I would first pre-train the network with step 4, then try step 5 and train long enough to let time to the network to go below the margin value.

[Hjy0801]

Thanks for your explanation,. I will do the step 4 and then do the step 5. Have a good week!