gegewen / nested-fno

Make dataloader

• step 1: download meta data from google drive and put them into Nested_FNO/ECLIPSE/meta_data
• step 2: run following code to convert .npy file into .pt files in dataset folder

cd data_config
bash file_config.sh
cd ..

• step 3: run python3 save_data_loader.py to create DATA_LOADER_DICT.pth

Training procedure

• step 1: train each models seperately using the following code. Each model requires an NVIDIA A100 GPU.

python3 train_FNO4D_DP_GLOBAL.py
python3 train_FNO4D_DP_LGR.py LGR1
python3 train_FNO4D_DP_LGR.py LGR2
python3 train_FNO4D_DP_LGR.py LGR3
python3 train_FNO4D_DP_LGR.py LGR4
python3 train_FNO4D_SG_LGR.py LGR1
python3 train_FNO4D_SG_LGR.py LGR2
python3 train_FNO4D_SG_LGR.py LGR3
python3 train_FNO4D_SG_LGR.py LGR4

• step 2: monitor training and validation loss with tensorboard

tensorboard --logdir=logs --port=6007 --host=xxxxxx

Finetune procedure

As discussed in the paper, we finetuned dP_LGR1, dP_LGR4, SG_LGR1, SG_LGR1 models with a random instance of pre-generated error.