precip_SR_demo

simple precipitation super-resolution demo in Pytorch-Lightning

Usage

set up env

critical packages: rootutils, wandb, omegaconf, torch, pytorch_lightning

set up data

raw & interim data is NOT provided due to our license, please download at 国家气象科学数据中心.

after obtaining the raw data, execute files in src/raw2dataset in numerical order, this will calculate pentad mean anomalies (i.e., deviation from grid annual cycle).

if you are looking for daily values rather than x-day-mean anomalies, modify line 79 in step3.

set up model

only simple UNet is implemented in this public version, and this will definitely lead to undesirable results.

but you can easily try other available SR models e.g. SRGAN, RCAN, etc.

just add some model to src/models.py

start training

modify config in src/config.yaml, or in cmd line, e.g., python train.py plmodule_config.model_name='RCAN'

project file tree

 .
 ├── data
 │   ├── dataset
 │   │   ├── CLDAS.meta.5d.csv
 │   │   ├── CLDAS.pre.1040x1600.5d.dat
 │   │   └── CLDAS.pre.260x400.LR.5d.dat
 │   ├── interim
 │   │   └── DOY
 │   │       ├── CLDAS_pre_0101_2008-2017.nc
            ...
 │   └── raw
 │       ├── 2008
 │       │   ├── 01
            ...
 ├── dependency
 │   ├── coords
 │   │   └── CLDAS.coords.npz
 │   └── stats
 │       ├── clim.CLDAS.pre.nc
 │       ├── clim.CLDAS.pre.smooth.nc
 │       └── CONST.CLDAS.pre.5d.txt
 ├── notebooks
 │   ├── check_whole_field.ipynb
 │   ├── read_patched_dataset.ipynb
 │   └── smooth_annual_cycle.ipynb
 └── src
     ├── raw2dataset
     │   ├── step1.concat_data.py
     │   ├── step2.save_clim.py
     │   ├── step3.to_dataset.py
     │   ├── step4.make_LR.py
     │   ├── step5.get_transform_stats.py
     │   └── utils
     │       └── nc_handler.py
     └── train
         ├── callbacks.py
         ├── config.yaml
         ├── dataset.py
         ├── models.py
         ├── plmodules.py
         └── train.py