Model factory

[AleShi94]

Possible solution to issue #286

1. model_factory and model_factory_from_env can take external neural network.
2. model_factory and model_factory_from_env can load a nn from file
3. model_factory and model_factory_from_env can load a checkpoint for nn
4. model_factory_from_env runs the check that the external nn satisfies requirements of the environment and model_configs

All new features are tested in test_torch_training.py file

[KohlerHECTOR]

I am not sure I understand the if/else structure of the model factory. What happens if both a filename and nn.module instance are passed to the model factory function ?

[AleShi94]

I am not sure I understand the if/else structure of the model factory. What happens if both a filename and nn.module instance are passed to the model factory function ?

well, technically you can do it so that you take architecture from nn.Module and from the file you can take model weights. In that case it will load weights from the file and assign them to the architecture

But to be honest I am not 100% sure if it is the best way to code stuff, but the idea is that it should be flexible enough:
either you use available tools to build a nn (like it was done before), or you can propose your own architecture through net argument or you can upload your model or checkpoint of the model through a file, which should allow transfer learning

[KohlerHECTOR]

Your idea is very clear and cool. What about only taking a nn module instance as input ? And providing the user with a simple pol net val net and a model factory ? So it is the user's task to build its networks the way he/she likes?
Otherwise: lgtm :)

[JulienT01]

I just added two one little comment.

And you need to run your tests (test_torch_training.py) inside functions that start with 'test_' (previous error, which needs to be corrected).

Otherwise it looks good to me. 👍