Mohammad Babaeizadeh

Implementing the LSTM version without lots of code change depends on how long the sequences of training data should be. If the sequences are as long as TMAX frames (which...

@etienne87 you are correct. But please look at [here](https://github.com/NVlabs/GA3C/blob/master/ga3c/ThreadTrainer.py#L50-L52). What Trainer receives is in (N, T, C, H, W) format but it merges the T dimension to have data in...

@etienne87 I'm not sure if I understand your first question about reset_state correctly. Can you please provide more details? About having separate graphs, there are different ways of implementing the...

@developeralgo8888 we don't have an eta yet but working on it. As @ifrosio mentioned, a naive multi-GPU implementing does not improve the converge rate and may cause instabilities. A naive...

This is a good feature to add but in our experiments it looks unnecessary. Usually the agent figures out the importance of pressing the fire button quickly enough and at...

That's an interesting observation. I've tested the code on a Maxwell TITAN X myself and didn't observe such behavior. Can you please share the version of your libraries (python, TensorFlow,...

@ifrosio that's a very good point. @developeralgo8888 please try with `DYNAMIC_SETTINGS=False`

Related to #4