pre-trained model, freeze-backbone etc..

[yerin723]

I'm labeling a total of 198 images that are 9920
We are trying to train the db_resnet50-CRNN model in addition to the existing pre-trained model.
Fine-tuning is being attempted on the list for model performance and the performance is that I like it.
We are adjusting the list of learning rates, batch sizes, epochs, Weight_dacay, workers, etc.
Are there clear values?

Also, if you think you have very small data compared to existing pre-trained models, you can use Freeze-backbone.
I am using method.
Is it better to keep the convolutional base fixed and learn a new opposite classifier? Or is it better to keep the convolutional base fixed and learn only the classifier?
If there is an exception, please let me know how to use --freeze_backbone. Just type --freeze_backbone?