VMarsocci · yurujaja · Oct 16, 2024 · Oct 11, 2024 · Oct 11, 2024 · Oct 11, 2024
diff --git a/.github/CONTRIBUTING.md b/.github/CONTRIBUTING.md
@@ -177,16 +177,16 @@ We have designed the repo to allow for using your own datasets with minimal effo
 1. **Implement a Dataset Class**:
 
    - In the `pangaea/datasets/` directory, create a new Python file named after your dataset (e.g., `my_dataset.py`).
-   - Implement a class that inherits from `GeoFMDataset`. You can check it in `pangaea/datasets/base.py`.
+   - Implement a class that inherits from `RawGeoFMDataset`. You can check it in `pangaea/datasets/base.py`.
    - Be sure that your dataset is instantiated with all the required parameters from the `GeoFMDataset`. You can also add new parameters.
    - Implement the required methods: `__init__`, `__len__`, `__getitem__`, and `download` (if applicable, otherwise a `NotImplementedError is raised`).
    - **Example**:
 
      ```python
      import torch
-     from pangaea.datasets.base import GeoFMDataset
+     from pangaea.datasets.base import RawGeoFMDataset
 
-     class MyDataset(GeoFMDataset):
+     class MyDataset(RawGeoFMDataset):
           def __init__(
              self,
              split: str,
@@ -236,6 +236,25 @@ We have designed the repo to allow for using your own datasets with minimal effo
              return len(self.file_list)
 
          def __getitem__(self, index):
+            """Returns the i-th item of the dataset.
+
+            Args:
+                i (int): index of the item
+
+            Raises:
+                NotImplementedError: raise if the method is not implemented
+
+            Returns:
+                dict[str, torch.Tensor | dict[str, torch.Tensor]]: output dictionary follwing the format
+                {"image":
+                    {
+                    "optical": torch.Tensor of shape (C T H W) (where T=1 if single-temporal dataset),
+                     "sar": torch.Tensor of shape (C T H W) (where T=1 if single-temporal dataset),
+                     },
+                "target": torch.Tensor of shape (H W) of type torch.int64 for segmentation, torch.float for
+                regression datasets.,
+                 "metadata": dict}.
+            """
              # Load your data and labels here
              image = ...  # Load image
              target = ...  # Load target label or mask

diff --git a/.gitignore b/.gitignore
@@ -1,13 +1,17 @@
 **/__pycache__/
 *.egg-info
 
+aaa_mystuff/
+work-dir/
 outputs/
 old_files/
 pretrained/
 pretrained_models/
 work-dir/
 wandb/
 data/*
+data
+pretrained_models
 # Scenes for mados tiny
 # Wow gitignore pattern mathcing is a mess
 !data/mados

diff --git a/configs/dataset/dynamicen.yaml b/configs/dataset/dynamicen.yaml
@@ -26,6 +26,14 @@ distribution:
   - 0.28
   - 0.08
 
+sample_dates:
+  - 1
+  - 5
+  - 10
+  - 15
+  - 20
+  - 25
+
 # data stats
 bands:
   optical:
@@ -36,15 +44,15 @@ bands:
 
 data_mean:
   optical:
-    - 1042.59240722656
-    - 915.618408203125
     - 671.260559082031
+    - 915.618408203125
+    - 1042.59240722656
     - 2605.20922851562
 data_std:
   optical:
-    - 957.958435058593
-    - 715.548767089843
     - 596.943908691406
+    - 715.548767089843
+    - 957.958435058593
     - 1059.90319824218
 
 data_min:

diff --git a/configs/dataset/pastis.yaml b/configs/dataset/pastis.yaml
@@ -10,8 +10,8 @@ multi_modal: True
 #limited_label: False
 
 # classes
-ignore_index: 0
-num_classes: 19
+ignore_index: 19
+num_classes: 20
 classes: 
   - Background
   - Meadow
@@ -32,7 +32,7 @@ classes:
   - Orchard
   - Mixed Cereal
   - Sorghum
-  #- Void Label
+  - Void Label
 distribution:
   - 0.00000
   - 0.25675
@@ -53,7 +53,7 @@ distribution:
   - 0.02460
   - 0.00696
   - 0.00580
-  #- 0.29476
+  - 0.29476
 
 bands:
   optical:

diff --git a/configs/encoder/croma_joint.yaml b/configs/encoder/croma_joint.yaml
@@ -28,3 +28,5 @@ output_layers:
   - 5
   - 7
   - 11
+
+output_dim: 1024
diff --git a/configs/encoder/croma_optical.yaml b/configs/encoder/croma_optical.yaml
@@ -24,3 +24,5 @@ output_layers:
   - 5
   - 7
   - 11
+
+output_dim: 1024
diff --git a/configs/encoder/croma_sar.yaml b/configs/encoder/croma_sar.yaml
@@ -14,4 +14,6 @@ output_layers:
   - 3
   - 5
   - 7
-  - 11
+  - 11
+
+output_dim: 1024
diff --git a/configs/encoder/dofa.yaml b/configs/encoder/dofa.yaml
@@ -41,3 +41,5 @@ output_layers:
   - 5
   - 7
   - 11
+
+output_dim: 768
diff --git a/configs/encoder/gfmswin.yaml b/configs/encoder/gfmswin.yaml
@@ -7,7 +7,6 @@ in_chans: 3
 t_patch_size: 3
 depth: 12
 embed_dim: 128
-output_dim: 1024
 img_size: 192    # fixed to 192 to avoid interpolation in checkpoints which leads to drop in performance
 depths: [ 2, 2, 18, 2 ]  
 num_heads: [ 4, 8, 16, 32 ]
@@ -28,3 +27,8 @@ output_layers:
   - 2
   - 3
 
+output_dim:
+  - 256
+  - 512
+  - 1024
+  - 1024
diff --git a/configs/encoder/prithvi.yaml b/configs/encoder/prithvi.yaml
@@ -27,3 +27,5 @@ output_layers:
   - 5
   - 7
   - 11
+
+output_dim: 768
diff --git a/configs/encoder/remoteclip.yaml b/configs/encoder/remoteclip.yaml
@@ -20,4 +20,6 @@ output_layers:
   - 3
   - 5
   - 7
-  - 11
+  - 11
+
+output_dim: 768
diff --git a/configs/encoder/satlasnet_mi.yaml b/configs/encoder/satlasnet_mi.yaml
@@ -5,7 +5,6 @@ download_url: null
 model_identifier: Sentinel2_SwinB_MI_MS # Multi-Image Multi-Spectral 
 fpn: False
 input_size: 128
-output_dim: 1024
 
 input_bands:
   optical:
@@ -23,3 +22,15 @@ input_bands:
     - B8
     - B11
     - B12
+
+output_layers:
+  - 0
+  - 1
+  - 2
+  - 3
+
+output_dim:
+  - 128
+  - 256
+  - 512
+  - 1024
diff --git a/configs/encoder/satlasnet_si.yaml b/configs/encoder/satlasnet_si.yaml
@@ -5,7 +5,6 @@ download_url: null
 model_identifier: Sentinel2_SwinB_SI_MS  # Single Image Multi-Spectral 
 fpn: False
 input_size: 128
-output_dim: 1024
 
 input_bands:
   optical:
@@ -23,3 +22,15 @@ input_bands:
     - B8
     - B11
     - B12
+
+output_layers:
+  - 0
+  - 1
+  - 2
+  - 3
+
+output_dim:
+  - 128
+  - 256
+  - 512
+  - 1024
diff --git a/configs/encoder/scalemae.yaml b/configs/encoder/scalemae.yaml
@@ -25,3 +25,5 @@ output_layers:
   - 11
   - 15
   - 23
+
+output_dim: 1024
diff --git a/configs/encoder/spectralgpt.yaml b/configs/encoder/spectralgpt.yaml
@@ -3,7 +3,7 @@ encoder_weights: ./pretrained_models/SpectralGPT+.pth
 download_url: https://zenodo.org/records/8412455/files/SpectralGPT+.pth
 
 input_size: 128
-output_dim: 3072   # 768 * (in_chans / t_patch_size)
+
 in_chans: 12    # number of spectral bands
 t_patch_size: 3
 depth: 12
@@ -34,4 +34,5 @@ output_layers:
   - 7
   - 11
 
+output_dim: 3072   # 768 * (in_chans / t_patch_size)
 
diff --git a/configs/encoder/ssl4eo_data2vec.yaml b/configs/encoder/ssl4eo_data2vec.yaml
@@ -31,3 +31,5 @@ output_layers:
   - 5
   - 7
   - 11
+
+output_dim: 384
diff --git a/configs/encoder/ssl4eo_dino.yaml b/configs/encoder/ssl4eo_dino.yaml
@@ -31,3 +31,5 @@ output_layers:
   - 5
   - 7
   - 11
+
+output_dim: 384
diff --git a/configs/encoder/ssl4eo_mae_optical.yaml b/configs/encoder/ssl4eo_mae_optical.yaml
@@ -31,4 +31,6 @@ output_layers:
   - 3
   - 5
   - 7
-  - 11
+  - 11
+
+output_dim: 384
diff --git a/configs/encoder/ssl4eo_mae_sar.yaml b/configs/encoder/ssl4eo_mae_sar.yaml
@@ -19,4 +19,6 @@ output_layers:
   - 3
   - 5
   - 7
-  - 11
+  - 11
+
+output_dim: 384
diff --git a/configs/encoder/ssl4eo_moco.yaml b/configs/encoder/ssl4eo_moco.yaml
@@ -31,3 +31,5 @@ output_layers:
   - 5
   - 7
   - 11
+
+output_dim: 384
diff --git a/configs/encoder/unet_encoder.yaml b/configs/encoder/unet_encoder.yaml
@@ -6,3 +6,9 @@ topology: [64, 128, 256, 512,]
 
 input_bands: ${dataset.bands}
 
+output_dim:
+  - 64
+  - 128
+  - 256
+  - 512
+
diff --git a/configs/preprocessing/reg_default.yaml b/configs/preprocessing/reg_default.yaml
@@ -1,35 +1,21 @@
-train: 
-  - _target_: pangaea.engine.data_preprocessor.RegPreprocessor
-    # overwritten in run
-    dataset: null
-    encoder: null
+train:
+    _target_: pangaea.engine.data_preprocessor.Preprocessor
+    preprocessor_cfg:
+        - _target_: pangaea.engine.data_preprocessor.RandomCropToEncoder
+        - _target_: pangaea.engine.data_preprocessor.BandFilter
+        - _target_: pangaea.engine.data_preprocessor.NormalizeMeanStd
+        - _target_: pangaea.engine.data_preprocessor.BandPadding
 
-  - _target_: pangaea.engine.data_preprocessor.NormalizeMinMax
-    # overwritten in run
-    dataset: null
-    encoder: null
-    data_min: -1
-    data_max: 1
+val:
+    _target_: pangaea.engine.data_preprocessor.Preprocessor
+    preprocessor_cfg:
+        - _target_: pangaea.engine.data_preprocessor.BandFilter
+        - _target_: pangaea.engine.data_preprocessor.NormalizeMeanStd
+        - _target_: pangaea.engine.data_preprocessor.BandPadding
 
-  - _target_: pangaea.engine.data_preprocessor.RandomCropToEncoder
-    # overwritten in run
-    dataset: null
-    encoder: null
-
-test: 
-  - _target_: pangaea.engine.data_preprocessor.RegPreprocessor
-    # overwritten in run
-    dataset: null
-    encoder: null
-
-  - _target_: pangaea.engine.data_preprocessor.NormalizeMinMax
-    # overwritten in run
-    dataset: null
-    encoder: null
-    data_min: -1
-    data_max: 1
-
-  - _target_: pangaea.engine.data_preprocessor.Tile
-    # overwritten in run
-    dataset: null
-    encoder: null
+test:
+    _target_: pangaea.engine.data_preprocessor.Preprocessor
+    preprocessor_cfg:
+        - _target_: pangaea.engine.data_preprocessor.BandFilter
+        - _target_: pangaea.engine.data_preprocessor.NormalizeMeanStd
+        - _target_: pangaea.engine.data_preprocessor.BandPadding
-Original file line number
+Diff line change
@@ Expand Up / @@ -28,3 +28,5 @@ output_layers: @@
       - 5
       - 7
       - 11
+    output_dim: 1024