Merge pull request #14 from instadeepai/13-update-instageo-model-to-c…

…onform-with-new-instageo-data-pipeline

fix: update instageo-model

instageo/model/README.md

@@ -38,7 +38,7 @@ pip install . #run from instageo root
     - `learning_rate`: Initial learning rate.
     - `num_epochs`: Number of training epochs.
     - `batch_size`: Batch size for training and validation.
-    - `mode`: Select one of training or evaluation mode.
+    - `mode`: Select training, evaluation or stats mode.
 See `configs/config.yaml` for more.
 
 2. **Dataset Preparation:** Prepare your geospatial data using the InstaGeo Chip Creator or similar and place it in the specified `root_dir`. Ensure that the csv file for each dataset has `Input` and `Label` columns corresponding to the path of the image and label relative to the `root_dir`. Additionally, ensure the data is compatible with `InstaGeoDataset`
@@ -221,5 +221,6 @@ When the saved checkpoint is evaluated on the test set, you should have results
 
 ## Customization
 
-- Modify the `bands`, `mean`, and `std` lists in `configs/config.yaml` to match your dataset's characteristics.
+- Use the `stats` mode to compute the `mean`, and `std` lists of your dataset.
+- Modify the `bands`, `mean`, and `std` lists in `configs/config.yaml` to match your dataset's characteristics and improve its normalization.
 - Implement additional data augmentation strategies in `process_and_augment`.

instageo/model/configs/locust.yaml

@@ -21,13 +21,29 @@ model:
 dataloader:
   # 3*(Blue, Green, Red, Narrow NIR, SWIR 1, SWIR 2)
   bands: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]
-  mean: [0.14245495, 0.13921481, 0.12434631, 0.31420089, 0.20743526, 0.12046503]
-  std: [0.04036231, 0.04186983, 0.05267646, 0.0822221, 0.06834774, 0.05294205]
+  mean:
+    [
+      623.2724609375,
+      1247.657958984375,
+      1772.24169921875,
+      2371.256103515625,
+      2862.867431640625,
+      2357.759765625,
+    ]
+  std:
+    [
+      2182.050048828125,
+      2248.420654296875,
+      2302.53515625,
+      2372.204345703125,
+      2398.52685546875,
+      2292.96435546875,
+    ]
   img_size: 224
   temporal_dim: 3
-  replace_label: null
+  replace_label: [-9999, -1]
   reduce_to_zero: False
-  no_data_value: -1
+  no_data_value: -9999
   constant_multiplier: 1.0
 
 test:

instageo/model/dataloader.py

@@ -244,14 +244,6 @@ def get_raster_data(
             data = src.read()
     if (not is_label) and bands:
         data = data[bands, ...]
-    # For some reasons, some few HLS tiles are not scaled in v2.0.
-    # In the following lines, we find and scale them
-    bands = []
-    for band in data:
-        if band.max() > 10:
-            band *= 0.0001
-        bands.append(band)
-    data = np.stack(bands, axis=0)
     return data
 
 

instageo/model/model.py

@@ -160,7 +160,12 @@ def __init__(
         if freeze_backbone:
             for param in model.parameters():
                 param.requires_grad = False
-        _ = model.load_state_dict(checkpoint, strict=False)
+        filtered_checkpoint_state_dict = {
+            key[len("encoder.") :]: value
+            for key, value in checkpoint.items()
+            if key.startswith("encoder.")
+        }
+        _ = model.load_state_dict(filtered_checkpoint_state_dict)
 
         self.prithvi_100M_backbone = model
 

instageo/model/run.py

@@ -418,6 +418,38 @@ def compute_metrics(
         }
 
 
+def compute_mean_std(data_loader: DataLoader) -> Tuple[List[float], List[float]]:
+    """Compute the mean and standard deviation of a dataset.
+
+    Args:
+        data_loader (DataLoader): PyTorch DataLoader.
+
+    Returns:
+        mean (list): List of means for each channel.
+        std (list): List of standard deviations for each channel.
+    """
+    mean = 0.0
+    var = 0.0
+    nb_samples = 0
+
+    for data, _ in data_loader:
+        # Reshape data to (B, C, T*H*W)
+        batch_samples = data.size(0)
+        data = data.view(batch_samples, data.size(1), -1)
+
+        nb_samples += batch_samples
+
+        # Sum over batch, height and width
+        mean += data.mean(2).sum(0)
+
+        var += data.var(2, unbiased=False).sum(0)
+
+    mean /= nb_samples
+    var /= nb_samples
+    std = torch.sqrt(var)
+    return mean.tolist(), std.tolist()  # type:ignore
+
+
 @hydra.main(config_path="configs", version_base=None, config_name="config")
 def main(cfg: DictConfig) -> None:
     """Runner Entry Point.
@@ -446,6 +478,34 @@ def main(cfg: DictConfig) -> None:
     test_filepath = cfg.test_filepath
     checkpoint_path = cfg.checkpoint_path
 
+    if cfg.mode == "stats":
+        train_dataset = InstaGeoDataset(
+            filename=train_filepath,
+            input_root=root_dir,
+            preprocess_func=partial(
+                process_and_augment,
+                mean=[0] * len(MEAN),
+                std=[1] * len(STD),
+                temporal_size=TEMPORAL_SIZE,
+                im_size=IM_SIZE,
+            ),
+            bands=BANDS,
+            replace_label=cfg.dataloader.replace_label,
+            reduce_to_zero=cfg.dataloader.reduce_to_zero,
+            no_data_value=cfg.dataloader.no_data_value,
+            constant_multiplier=cfg.dataloader.constant_multiplier,
+        )
+        train_loader = create_dataloader(
+            train_dataset,
+            batch_size=batch_size,
+            shuffle=True,
+            num_workers=1,
+        )
+        mean, std = compute_mean_std(train_loader)
+        print(mean)
+        print(std)
+        exit(0)
+
     if cfg.mode == "train":
         check_required_flags(["root_dir", "train_filepath", "valid_filepath"], cfg)
         train_dataset = InstaGeoDataset(