HDF5Output for the logger

rlworkgroup / dowel

A little logger for machine learning research

MIT License

32 stars 37 forks source link

HDF5Output for the logger #10

Open amolchanov86 opened 5 years ago

amolchanov86 commented 5 years ago

Hi guys, I am currently in need of recording trajectories of experiments, hence, I will add my own hdf5 logger. Hence the feature request to have the capability to log HDF5 files with data in the future.

Justification: The current params.pkl is not a very convenient way to store trajectories (and other heavy statistics/data) for a few reasons:

pkl files usually quite large, hdf5 files are more compact
current params.pkl requires running a session and using joblib. I need more framework-independent storage (for example, you can imagine if I would like to use Matlab later to load trajectories and do some postprocessing for the paper).

ryanjulian commented 5 years ago

Can you provide examples of what you want to send to HDF5?

Params? Trajectories? Training progress?

amolchanov86 commented 5 years ago

At the moment I am saving trajectories (i.e. "paths"). Btw, I have already written a small, but more or less generic library for saving (and loading from) dictionaries into HDF5, but I haven't wrapped it inside of the logger. I also need a capability to provide a single path for a bunch of experiments to store all paths in a single h5, hence it is kept separate for now. But I can share this lib if someone is willing to add the feature to the garage.

amolchanov86 commented 5 years ago

It is also convenient to store class arguments inside of the HDF5, hence the paths and the parameters that were used to generate them are kept in the same place. RIght now I am using the following structure inside HDF5 due to the specifics of my project:

# teacher_info/
#   - [teacher_indx]: 
# traj_data/ 
#   - [teacher_indx] * [iter_indx] * traj_data

For example _trajdata/0000/0000 refers to data of the first run (teacher) in the first iteration Where by "teacher" I mean the optimization run. Iteration is just a ppo policy iterations. Teacher info contains parameters associated with the run.

Hope it helps.

ryanjulian commented 5 years ago

@pelillian is rewriting the logger now and it seems like this could be a good addition.

it's meant to be a single frontend with pluggable backends. this could be a great project to see if the plugging system works nicely.

@krzentner might be interested in this issue.