Yet-another-efficientdet-pytorch: training custom dataset problem

[tdn670000]

I'm trying to detect car plate number(1 class), but cannot get good quality
using repo version of 14.04.2020 16:13 utc

I compared formats of dataloader outputs with your tutorial, its ok
set force_input_size = None
current training status:
Val. Epoch: 186/500. Classification loss: 0.54055. Regression loss: 0.05754. Total loss: 0.59809
Step: 34772. Epoch: 187/500. Iteration: 178/185. Cls loss: 0.36639. Reg loss: 0.04083. Total loss: 0.40722: 96%
178/185 [04:41<00:11, 1.59s/it]

learning rate = 1e-4
total loss isn't decreasing about last 30 epochs. and not increasing, so it's not look like it ovefitting

prediction(max confidence of bbox = 0.092):

what can i do?
what value of loss is good?

[zylo117]

overfitting.

try --head_only True

[tdn670000]

I trained with head_only True

I understood, it was underfitting
So, I trained with "lr"=1e-3 and "head_only"=True and "load_weights" d0 from repo first
then lr=1e-3 and head_only False
then lr=1e-4

and it's result(confidence=0.974):

Thank you

[zylo117]

good job!

[xkp793003821]

@tdn670000
good job,
can you share the final class_loss and reg_loss in your experiment when you get a good test result
thanks

[tdn670000]

But I have a new problem, my model cannot predict small objects, so I think, classification loss should be less

[zylo117]

@tdn670000
yes, the weights are overfitting in the final 50k steps. you should fix it and then test again.

[rm2886]

@tdn670000 @zylo117 can you please tell me how to view tensorboard page.
I tried but always get a white page.

Thanks

[zylo117]

tensorboard --logdir /path/to/the/tensorboard/log/folder/

[xkp793003821]

@tdn670000 @zylo117
Thanks for your work, I have get a good detection result
ahh

[mytk2012]

I trained with head_only True

I understood, it was underfitting
So, I trained with "lr"=1e-3 and "head_only"=True and "load_weights" d0 from repo first
then lr=1e-3 and head_only False
then lr=1e-4

and it's result(confidence=0.974):

Thank you

Can you tell me how to visual my results except loss？I am not sure if my test results are correct