DPOT

DPOT: Auto-Regressive Denoising Operator Transformer for Large-Scale PDE Pre-Training (ICML'2024)

Code for paper DPOT: Auto-Regressive Denoising Operator Transformer for Large-Scale PDE Pre-Training (ICML'2024). It pretrains neural operator transformers (from 7M to 1B) on multiple PDE datasets. Pre-trained weights could be found at https://huggingface.co/hzk17/DPOT.

Our pre-trained DPOT achieves the state-of-the-art performance on multiple PDE datasets and could be used for finetuning on different types of downstream PDE problems.

Usage

Pre-trained models

We have five pre-trained checkpoints of different sizes. Pre-trained weights are at https://huggingface.co/hzk17/DPOT.

| Size | Attention dim | MLP dim | Layers | Heads | Model size | | ------ | ------------- | ------- | ------ | ----- | ---------- | | Tiny | 512 | 512 | 4 | 4 | 7M | | Small | 1024 | 1024 | 6 | 8 | 30M | | Medium | 1024 | 4096 | 12 | 8 | 122M | | Large | 1536 | 6144 | 24 | 16 | 509M | | Huge | 2048 | 8092 | 27 | 8 | 1.03B |

Here is an example code of loading pre-trained model.

model = DPOTNet(img_size=128, patch_size=8, mixing_type='afno', in_channels=4, in_timesteps=10, out_timesteps=1, out_channels=4, normalize=False, embed_dim=512, modes=32, depth=4, n_blocks=4, mlp_ratio=1, out_layer_dim=32, n_cls=12)
model.load_state_dict(torch.load('model_Ti.pth')['model'])

Dataset Protocol

All datasets are stored using hdf5 format, containing data field. Some datasets are stored with individual hdf5 files, others are stored within a single hdf5 file.

In data_generation/preprocess.py, we have the script for preprocessing the datasets from each source. Download the original file from these sources and preprocess them to /data folder.

| Dataset | Link | | ------------- | ------------------------------------------------------------ | | FNO data | Here | | PDEBench data | Here | | PDEArena data | Here | | CFDbench data | Here |

In utils/make_master_file.py , we have all dataset configurations. When new datasets are merged, you should add a configuration dict. It stores all relative paths so that you could run on any places.

mkdir data

Single GPU Pre-training

Now we have a single GPU pretraining code script train_temporal.py, you could start it by

python train_temporal.py --model DPOT --train_paths ns2d_fno_1e-5 --test_paths ns2d_fno_1e-5 --gpu 0 

to start a training process.

Or you could start it by writing a configuration file in configs/ns2d.yaml and start it by automatically using free GPUs with

python trainer.py --config_file ns2d.yaml

Multiple GPU Pre-training

python parallel_trainer.py --config_file ns2d_parallel.yaml

Configuration file

Now I use yaml as the configuration file. You could specify parameters for args. If you want to run multiple tasks, you could move parameters into the tasks ,

model: DPOT
width: 512
tasks:
 lr: [0.001,0.0001]
 batch_size: [256, 32] 

This means that you start 2 tasks if you submit this configuration to trainer.py.

Requirement

Install the following packages via conda-forge

conda install pytorch==2.0.0 torchvision==0.15.0 torchaudio==2.0.0 pytorch-cuda=11.7 -c pytorch -c nvidia
conda install matplotlib scikit-learn scipy pandas h5py -c conda-forge
conda install timm einops tensorboard -c conda-forge

Code Structure

• README.md
• train_temporal.py: main code of single GPU pre-training auto-regressive model
• trainer.py: framework of auto scheduling training tasks for parameter tuning
• utils/
  • criterion.py: loss functions of relative error
  • griddataset.py: dataset of mixture of temporal uniform grid dataset
  • make_master_file.py: datasets config file
  • normalizer: normalization methods (#TODO: implement instance reversible norm)
  • optimizer: Adam/AdamW/Lamb optimizer supporting complex numbers
  • utilities.py: other auxiliary functions
• configs/: configuration files for pre-training or fine-tuning
• models/
  • dpot.py: DPOT model
  • fno.py: FNO with group normalization
  • mlp.py
• data_generation/: Some code for preprocessing data (ask hzk if you want to use them)
  • darcy/
  • ns2d/

Citation

If you use DPOT in your research, please use the following BibTeX entry.

@article{hao2024dpot,
  title={DPOT: Auto-Regressive Denoising Operator Transformer for Large-Scale PDE Pre-Training},
  author={Hao, Zhongkai and Su, Chang and Liu, Songming and Berner, Julius and Ying, Chengyang and Su, Hang and Anandkumar, Anima and Song, Jian and Zhu, Jun},
  journal={arXiv preprint arXiv:2403.03542},
  year={2024}
}