Topic of Masterarbeit_FAU_HHI: Discrete Data Generation with Trajectory Models

Contents

The current code consists of two parts: modified DDSM and modified DDPM.

Description

How to set up the environment

Use base_env_pytorch.def to set up the environment to get a .sif file with the same name.

How to run modified DDSM

1. Run MA_2024_prepsamples.sh to get a simulated diffusion path. In this .sh file, the presample_noise.py will be called. There are several parameters which can be set up:

   • -n, --num_samples, type=int
   • -c, --num_cat, type=int
   • -t, --num_time_steps, type=int
   • --speed_balance, type=bool, default=false
   • --max_time, type=float
   • --out_path, type=str
   • --boundary_mode, choices=['clamp', 'reflect_boundaries', 'reflection'], default='clamp'

   Example:

   presample_noise.py -n 50000 -c 9 -t 400 --max_time 1 --out_path sudoku/ --boundary_mode 'reflect_boundaries'

2. Run MA_2024_results_ddsm_channel_std.sh to check the quality of path simulation end distribution.

3. Run MA_2024_train.sh to train the model. The .pth files are saved after each 50 epochs. There are several parameters which can be set up:

   • -t, --num_time_steps, type=int, default=400
   • -spd, --speed_balance, choices=['s1', 'sab'], default='s1'
   • -bm, --boundary_mode, choices=['clamp', 'reflect_boundaries', 'reflection'], default='clamp'

   Example:

   train_sudoku_allinone.py -t 100 -spd 's1' -bm 'clamp'

4. Run MA_2024_eval_Epoch.sh to evaluate the trained model.

How to run modified DDPM

1. Run SH_sudoku_csv_to_pickle.sh to extract Sudoku data from a pre-randomly created Sudoku .csv file, or create random Sudoku during the training phase. (Preferably, extract data from a .csv file to save training time.)

2. Run SH_ddpm_train.sh to train the model. You can change the UNet size in ./ddpm/nets/unet.py.

3. Run SH_ddpm_predict.sh to evaluate the trained model.

Environment requirements

• Python 3.8+
• CUDA: 11.7.1-cudnn8-devel-ubuntu20.04