Add pipeline tag and library name

README.md

@@ -1,5 +1,7 @@
 ---
 license: mit
+pipeline_tag: image-feature-extraction
+library_name: torch
 ---
 
 # AnySat: An Earth Observation Model for Any Resolutions, Scales, and Modalities
@@ -62,7 +64,7 @@ Arrange your data in a dictionary with any of the following keys:
 | l8            | Time series tensor |BxTx11xHxW                                           | B8, B1, B2, B3, B4, B5, B6, B7, B9, B10, B11 | 10m     |
 | modis         | Time series tensor |BxTx7xHxW                                            | B1, B2, B3, B4, B5, B6, B7                | 250m       |
    
-Note that time series requires a `_dates` companion tensor containing the day of the year: 01/01 = 0, 31/12=364.
+Note that time series requires a `_dates` companion tensor containing the day of the year: 01/01 = 0, 31/12=364. The data also needs to be normalized (data -mean) /std as input to our model.
 
 **Example Input** for a tile of 60x60m and a batch size of B:
 
@@ -84,16 +86,19 @@ Decide on:
   - `'tile'`: Single vector per tile
   - `'patch'`: A vector per patch
   - `'dense'`: A vector per sub-patch
-  - `'all'`: a tuple with all three outputs
- 
+  - `'all'`: A vector per patch with class token at first position
+
+⚠️ For segmentation tasks, use 'dense' argument!
+
 The sub patches are `1x1` pixels for time series and `10x10` pixels for VHR images. If using `output='dense'`, specify the `output_modality`.
 
 Example use:
 ```python
-features = AnySat(data, scale=10, output='tile') #tensor of size [D,]
-features = AnySat(data, scale=10, output='patch') #tensor of size [D,6,6]
-features = AnySat(data, scale=20, output='patch') #tensor of size [D,3,3]
-features = AnySat(data, scale=20, output='dense', output_modality='aerial') #tensor of size [D,30,30]
+features = AnySat(data, patch_size=10, output='tile') #tensor of size [B, D,]
+features = AnySat(data, patch_size=10, output='patch') #tensor of size [B, D, 6, 6]
+features = AnySat(data, patch_size=20, output='patch') #tensor of size [B, D, 3, 3]
+features = AnySat(data, patch_size=20, output='dense', output_modality='aerial') #tensor of size [B, 2*D, 30, 30]
+features = AnySat(data, patch_size=20, output='dense', output_modality='s2') #tensor of size [B, 2*D, 6, 6]
 ```
 **Explanation for the size of the dense map:** `d=10` for 'aerial' which has a 0.2m resolution, the sub-patches are 2x2 m.
 
@@ -149,7 +154,7 @@ python src/train.py exp=BraDD_AnySAT_LP
 Our implementation already supports 9 datasets:
 
 <p align="center">
-  <img src="https://cdn-uploads.huggingface.co/production/uploads/662b7fba68ed7bbf40bfb0df/nGGz8kiDdeTJqIPSdrlSz.png" alt="AnySat Datasets" width="500">
+  <img src=".media/datasets.png" alt="AnySat Datasets" width="500">
 </p>
 
 ## GeoPlex Datasets
@@ -216,4 +221,10 @@ Please use the following bibtex:
   journal={arXiv preprint arXiv:2412.14123},
   year={2024}
 }
-```
+```
+
+# Acknowledgements
+- The code is conducted on the same base as [OmniSat](https://github.com/gastruc/OmniSat)
+- The JEPA implementation comes from [JEPA](https://github.com/facebookresearch/ijepa)
+- The code from Pangaea datasets comes from [Pangaea](https://github.com/VMarsocci/pangaea-bench)
+<br>