The default_curl_v3.ipynb notebook includes our current implementation of the CURL architecture (namely, the image encoder) along with our own data augmentations.
The resnet_pretrained.ipynb notebook uses the same RL architecture as above but uses a pretrained ImageNet ResNet model as the image encoder, which is used to contrast against the CURL encoder.
The curl+some_state-cartpole.ipnyb notebook is similar to the default CURL implementation except some velocity state is added to the image encodings to experiment with whether that is functionally equivalent to frame stacking.
The frame stacked variants are similar to the above but use two frames per observation instead of with the hopes of being able to natively encode velocity information between timesteps.
The pixel_baseline.ipynb notebook is a DQN implementation without encoding that is used a baseline of learning off pixels directly.
Run conda env create -f environment.yml and conda activate cs230proj. Then, run jupyter notebook to open the Jupyter Notebook webpage. Open the different notebooks and enjoy!
Runs on Linux only. May require xvfb package on headless systems.