Is the model_main_tf2.py just for academic purposes or could be used on real scenarios?

General Discussion
, , ,
#1

I tried a simple object training with github/tensorflow/models/blob/master/research/object_detection/model_main_tf2.py using several pre-trained models and my own images, but after hours of training the learning_rate is zero and loop continues for ever.

Someone advice my to review or reconfigure the training parameters like: reduceLRonPlateau

But this param is not found on model_main_tf2.py or other files in the official repository

If found this parameter on Keras, for example: tf.keras.callbacks.ReduceLROnPlateau  |  TensorFlow Core v2.8.0

My question are:

  • Is the model_main_tf2.py just for academic purposes or could be used on real scenarios?
  • If model_main_tf2.py where can I set or configure advanced training parameters pipeline.config?

Thank you so much

#3

Are you sure that is really “forever”?
What is see is that that training script has NUM_TRAIN_STEPS=10000:

If you want you can introduce an early_stopping policy:

#5

According to your link, early_stopping is on Keras. I don’t find this parameter on the classic models/model_main_tf2.py at master · tensorflow/models · GitHub

So, Can we say that model_main_tf2.py is just for academic purposes and Keras or others are the only option?

Thanks

#6

These models and scripts are under the research folder but you can always change some hyperparameters to adapt It to your pecific finetuning dataset (e.g. number of iteration, inital learning rate, etc.).

And a model refactoring Is going to be collected in: