Improves on the SeqGAN idea by adding more reinfocement learning and GAN techniques like:
- a replay buffer
- Consensus Optimization
- count-based exploration bonus
- Proximal Policy Optimization (was not found to help, but can be found in this commit)
- advantage normalization
If you wish to enable Consensus Optimization (via the --grad-reg option), you'll need
to patch PyTorch to allow forcing the use a
twice-differentiable RNN.
python3 main.py will run the project with the default options.
Output will be written to the run/ directory.
The em tool makes it really easy to twiddle hyperparameters by tracking changes to code (no need to make everything an option!).
Just run em run -g 0 exp_name with your desired options and you'll find a reproducable
snapshot in experiments/<exp_name>!
If you want to resume from a snapshot (perhaps with different options),
use em resume -g 0 exp_name ...
You can also fork an experiment and its changes using em fork, but the quick and dirty
solution is to run bash scripts/make_links.sh :)