Adapted from https://github.com/ehoogeboom/multinomial_diffusion
In the folder containing setup.py, run
pip install --user -e .
The --user option ensures the library will only be installed for your user.
The -e option makes it possible to modify the library, and modifications will be loaded on the fly.
You should now be able to use it.
Go to the folder bandit_diffusion:
cd bandit_diffusion
Run command to train the model:
python train.py --batch_size 32 --update_freq 1 --lr 0.01 --epochs 1000 --eval_every 2 --check_every 20 --diffusion_steps 1000 --transformer_depth 12 --transformer_heads 16 --transformer_local_heads 8 --gamma 0.99 --log_wandb False
Run command to sample new trajectories based on the trained model (the checkpoint is saved in ~/log/flow/bandit/multinomial_diffusion_v2/expdecay/YYYY-MM-DD_hh-nn-ss):
python eval_sample.py --length 512 --model "~/log/flow/bandit/multinomial_diffusion_v2/expdecay/YYYY-MM-DD_hh-nn-ss" --samples 16
The length should match the shape of samples (The (state-)action sequence length). Modify the file /datasets/data.py and /datasets/dataset_bandit.py to align with your data.