tobiapoppi / continual-learning-rl-project

continual-learning-RL-project

To prepare your environment you need to follow these instructions. Please, if you can, mind to the versions because it's easy to have incompatibilities.

1. create a new conda env

   • conda create --name rl python=3.8
   • conda activate rl

2. if you are on Windows AND you have a gpu and cuda installed (it's not mandatory!):

   for cuda <=11.7 (I have 11.2 and it's working fine):

   • pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117

   for cuda 11.8:

   • pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

3. if you don't have a gpu/cuda installed OR you are on Linux:

   • pip install torch torchvision

4. install other packets:

   • pip install ray gym[atari]==0.23.1 gym[accept-rom-license]==0.23.1 opencv-python atari-py pygame numpy

5. install avalanche 0.3.1 (note that you will need to change some files there, so I suggest you to install it manually with -e mode, and to modify the code as follows):

   • download this zip file: https://github.com/ContinualAI/avalanche/archive/refs/tags/v0.3.1.zip
   • unzip it and go with the terminal inside its folder
   • pip install -e .
   • go to /avalanche/training/plugins/ewc.py and substitute line 76 and 116 with this line: exp_counter = strategy.training_exp_counter

6. install avalanche-rl

   • pip install git+https://github.com/ContinualAI/avalanche-rl.git

To test if your env is working fine just activate the conda env and run:

python ewc_avalanche_load.py

You should see the emulator opening the pong game failing miserably (but you should at lease see it playing).