This reposity is holds the code needed to reproduce the results for my thesis. The Grad-CAM implementation used is based on https://github.com/kazuto1011/grad-cam-pytorch, the A3C implementation is based on https://github.com/ikostrikov/pytorch-a3c.
Some example analysis can be found here.
Note that it is still a work in progress, but eventually it should look like this:
python3 main.py --env-name ENV_NAME --gradcam_layer GCAM_LAYER
This command should run an episode of ENV_NAME using the pretrained models and collect Grad-CAM outputs for each state/action at layer GCAM_LAYER (which would default to features.elu4).
The models can be found in the pretrained folder. Each model has both a 'full' and a 'half' version (see paper).
| Full Agent Mean | Full Agent Variance | Half Agent Mean | Half Agent Variance | DeepMind | |
|---|---|---|---|---|---|
| Pong | 21.00 | 0.00 | 14.99 | 0.09 | 10.7 |
| BeamRider | 4659.04 | 1932.58 | 1597.40 | 1202.00 | 24622.2 |
| Seaquest | 1749.00 | 11.44 | N/A | N/A | 1326.1 |
Note that these models are based on amount of frames whereas DeepMind is based on 4 day training on 16 CPU cores, which makes comparing them hard.