Modelpublic

lschmidt/edsr-dsc

sourceHugging Facemitupdated 1y agoView on Hugging Face

0likes11downloads

Model Card

EDSR-DSC (4× Downscaling of Wind Velocities)

This model is a custom-trained version of the Enhanced Deep Super-Resolution (EDSR) model from the `super-image` library. It is adapted for downscaling of 2-channel ERA5 data (e.g., wind u and v components), by a factor of 4× (trained using COSMO-REA6 as high-resolution data).

🧠 Model Architecture

—Base: EDSR (Lim et al. 2017)
—Input channels: 2 (U & V components of wind speed)
—Output channels: 2
—Feature channels (`n_feats`): 64
—Residual blocks: 32
—Mean-shift normalization: Removed
—Upsampling: Enabled
—Scale factor: 4×

📦 Files in this Repository

File	Description
`config.json`	Configuration for the modified EDSR model
`pytorch_model_4x.pt`	Pretrained weights for 4× upscaling

🚀 How to Use

python

from super_image import EdsrModel, EdsrConfig
from huggingface_hub import hf_hub_download
import torch
import xarray as xr
import numpy as np

# load config 
config, _ = EdsrConfig.from_pretrained("lschmidt/edsr-dsc")

# load model & remove normalization
model = EdsrModel(config)
del model.sub_mean
del model.add_mean

# load pre-trained weights
state_dict_path = hf_hub_download(repo_id="lschmidt/edsr-dsc", filename="pytorch_model_4x.pt")
state_dict = torch.load(state_dict_path, map_location="cpu")
model.load_state_dict(state_dict, strict=False)

# create random input: must be a 4D tensor (B, C=2, H, W)
inputs = torch.randn(1, 2, 40, 40)  # replace with coarse wind velocity fields

# or use sample data
data_path = hf_hub_download(
    repo_id="lschmidt/edsr-dsc",
    filename="test_wind_velocities.nc",
    subfolder="test_data"  
)
ds = xr.open_dataset(data_path)
u = ds["u100"].values[0]
v = ds["v100"].values[0]
inputs = torch.from_numpy(np.stack([u, v], axis=0)).unsqueeze(0).float()  

# prediction
outputs = model(inputs)