Training result is abnormal without pre-trained weights.

[KevinRyu]

Hi,

Thanks for sharing your Yolo_v2 implementation using tf 2.0, first.
I trained this neural network after remarking "Load Pre-trained weights" code lines.
But both main and validation loss is not decreased at all. I got a loss graph looks like spike.
I think training is not proceeded correctly. Would you please give me an advice to solve this problem?
For example, I need to adjust hyper parameters, and so on.

[jmpap]

Hi,
I tried to train the model without pre-trained weights. Train loss decreases but validation loss increases. I am not sure it'is possible to train yolo v2 on our own dataset without pre-trained weights. According to the original article, the author explains that the pre-trained weights come from a classification model trained on ImageNet.

[KevinRyu]

Hi,

Thanks for your comment. I misunderstood code lines for loading pretrained weight (output for inference) that you have already been training with your data. I got it. I'll restore remarked code lines again and train it with custom data. By the way, you said train loss decreases. But I didn't see it. In my case, both train and validation loss is abnormal. I didn't understand your codes correctly. I will try again.

[jmpap]

Hi, if your train loss increases or is unstable, you can try a lower learning rate.

[KevinRyu]

OK! I will try. I learned many important points what I didn't know from your code. Your code is very clean and nice. Thank you very much. 2020년 4월 25일 (토) 오전 4:20, jmpap <notifications@github.com>님이 작성:

…

Hi, if your train loss increases or is unstable, you can try a lower learning rate. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#6 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AG43MBN7PWKJBOJ7YW7FTZTROHRA5ANCNFSM4MN5DZIA> .

[jmpap]

Thank your for your appreciation!