Transfer learning from a different model than the original one about leafnats HOT 3 CLOSED

After validation, you can find a folder nats_results/model/. This folder stores the best models. The next thing you need to do:

1. Go to model.py
   Change the modules that you want to transfer from self.train_models to self.base_models.
2. Go to main.py
   set --base_model_dir to the directory where you store your model files.

from leafnats.

Alright, I've done that. Now the questions that arise are:

1. In main.py, should I set the mode to transfer learning or using the pretrained parameters? (option --train_base_model)
2. Also in main.py, should I set --continue_training to true or false? From what I understand I'm generating a new model from a base one, so it should be false on the first execution (then, if I need to continue training I set it to true).
3. Should I be using pointer_generator_network or pointer_generator_network_trans? I have trained the model using pointer_generator_network, so when I try to use pointer_generator_network_trans I get an error message saying that embedding_base doesn't exist.

Thanks for the quick reply and for the good work done here, really.

from leafnats.

-train_base_model
true: if you want to also train those parameters.
false: they will be fixed during training.

-continue_training
most of the time, you should let it be true. In case you lost power, or your job is accidentally stopped, you can continue the training.

-you can use either of them. I suggest you use pointer-generator-network, since you can choose which part will be trained and which will be fixed.

from leafnats.