Pytorch IMPALA

A Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures (IMPALA) implemented in pytorch

Requirements

1. python 3.7+

$ sudo apt install python3.7

2. pytorch

$ pip install torch

3. tensorboard

$ pip install tensorboard

Usage

1. Edit hyperparameters in main.py

2. Train the model

$ python train.py

3. Logs will be collected in specified folder. You can use tensorboard to view them in a browser

$ tensorboard --logdir ./logs/

4. Test the model

$ python test.py

Example

$ python test.py -pp ./models/IMPALA_RacecarBulletEnv-v0_400.pt -hd 32 -en RacecarBulletEnv-v0 -ne 10 -el 1000 -ld ./logs/

References

1. [IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures by Espeholt, Soyer, Munos et al.] (https://arxiv.org/pdf/1802.01561.pdf)

TODO

• Fix OSError
• Add batched updates
• Add tensorboard logging
• Test performance
• Fix performance collapse issues
• Add comand line argument support