Find here pretrained model weights for the [Decision Transformer] (https://github.com/kzl/decision-transformer). Weights are available for 4 Atari games: Breakout, Pong, Qbert and Seaquest. Found in the checkpoints directory. We share models trained for one seed (123), whereas the paper contained weights for 3 random seeds.
Usage
git clone https://huggingface.co./edbeeching/decision_transformer_atari
conda env create -f conda_env.yml
Then, you can use the model like this:
from decision_transform_atari import GPTConfig, GPT
vocab_size = 4
block_size = 90
model_type = "reward_conditioned"
timesteps = 2654
mconf = GPTConfig(
vocab_size,
block_size,
n_layer=6,
n_head=8,
n_embd=128,
model_type=model_type,
max_timestep=timesteps,
)
model = GPT(mconf)
checkpoint_path = "checkpoints/Breakout_123.pth" # or Pong, Qbert, Seaquest
checkpoint = torch.load(checkpoint_path)
model.load_state_dict(checkpoint)