| # Equidiff | |
| - folder_name: the name of the folder | |
| - file_name: the name of your file | |
| ## Prepare data | |
| Use mimicgen to generate data. | |
| Use `EmbodiedBM/equidiff/combinehdf5.py` to combine data from multiple .hdf5 files if needed. | |
| Put hdf5 data at `EmbodiedBM/equidiff/data/robomimic/datasets` with the format [folder_name]/[file_name].hdf5 | |
| ## Convert data | |
| ```bash | |
| python equi_diffpo/scripts/robomimic_dataset_conversion.py -i data/robomimic/datasets/square_d2_test/demo.hdf5 -o data/robomimic/datasets/square_d2_test/demo_abs.hdf5 -n 12 | |
| ``` | |
| ## Train | |
| Use another CUDA device if 7 is currently in use. | |
| ```bash | |
| CUDA_VISIBLE_DEVICES=5 MUJOCO_GL=osmesa PYOPENGL_PLATFORM=osmesa HYDRA_FULL_ERROR=1 python train.py --config-name=train_sq2_5000 folder_name=square_d2_5000 file_name=demo n_demo=5000 | |
| ``` | |
| If you use another task than square_d2, you should change the task_name config by adding task_name=[task_name] | |
| ## Test | |
| Change the `ckpt_path` to the trained policy's weight's path in `EmbodiedBM/equidiff/equi_diffpo/config/test_sq2.yaml` | |
| If you use another task than square_d2, you should change the dataset config in test_sq2.yaml and download the corresponding dataset from [Huggingface](https://huggingface.co/datasets/amandlek/mimicgen_datasets/tree/main/core). | |
| ```bash | |
| python test.py | |
| ``` |