Create README.md

README.md

@@ -0,0 +1,47 @@
+---
+license: mit
+tags:
+  - image-super-resolution
+  - climate
+  - wind-power
+  - torch
+  - rcan
+library_name: super-image
+inference: false
+---
+
+# RCAN-DSC: Residual Channel Attention Network for Downscaling Wind Fields
+
+This model is a modified version of [RCAN](https://arxiv.org/abs/1807.02758), originally designed for image super-resolution, adapted for use in **climate downscaling** tasks. Specifically, it is configured to process **2-channel input data** representing wind vector components (u, v) at low spatial resolution and produce higher-resolution output fields.
+
+## 🧠 Model Description
+
+- Based on the original RCAN architecture from `super-image`.
+- `sub_mean` and `add_mean` normalization layers have been **removed**, as normalization of physical variables differs from image RGB workflows.
+- Supports **multi-channel inputs**, currently set up for **2-channel wind fields**.
+- Custom head and tail layers allow flexible channel dimensions and upscaling factors (e.g., ×2, ×4).
+- Optional partial weight loading from pretrained RCAN (trained on DIV2K) to initialize the main body.
+
+## 🛰️ Intended Use
+
+This model is intended for:
+- **Super-resolution of wind fields** in climate modeling applications.
+- **Downscaling wind speed or wind power-related variables** from coarse-resolution reanalysis or GCM output.
+- Research on generalizing deep learning architectures to geoscience and Earth system science domains.
+
+## 🧪 Example
+
+```python
+import torch
+from custom_rcan import RcanModelM  # Replace with your actual module
+from huggingface_hub import hf_hub_download
+
+# Load model
+model = RcanModelM(in_channels=2, scaling_factor=4)
+state_dict = torch.load(hf_hub_download(repo_id="your-username/rcan-dsc", filename="pytorch_model_4x.pt"))
+model.load_state_dict(state_dict, strict=False)
+
+# Dummy input: shape (B, C=2, H, W)
+x = torch.randn(1, 2, 64, 64)
+y = model(x)
+print(y.shape)  # (1, 2, 256, 256) for scale=4