Added visualisations for image augmentation

lerobot/common/datasets/transforms.py

@@ -47,11 +47,11 @@ class RandomSubsetApply(Transform):
 
 def make_transforms(cfg):
     image_transforms = []
-    if 'jit' in cfg.image_transform.list:
+    if 'colorjitter' in cfg.list:
-        image_transforms.append(v2.ColorJitter(brightness=cfg.colorjitter_range, contrast=cfg.colorjitter_range))
+        image_transforms.append(v2.ColorJitter(brightness=cfg.colorjitter_factor, contrast=cfg.colorjitter_factor))
-    if 'sharpness' in cfg.image_transform.list:
+    if 'sharpness' in cfg.list:
         image_transforms.append(v2.RandomAdjustSharpness(cfg.sharpness_factor, p=cfg.sharpness_p))
-    if 'blur' in cfg.image_transform.list:
+    if 'blur' in cfg.list:
         image_transforms.append(v2.RandomAdjustSharpness(cfg.blur_factor, p=cfg.blur_p))
 
-    return v2.Compose(RandomSubsetApply(image_transforms, n_subset=cfg.n_subset), v2.ToDtype(torch.float32, scale=True))
+    return v2.Compose([RandomSubsetApply(image_transforms, n_subset=cfg.n_subset), v2.ToDtype(torch.float32, scale=True)])

lerobot/configs/default.yaml

@@ -60,10 +60,9 @@ wandb:
 
 image_transform:
   enable: false
-  colorjittor_range: (0, 1)
+  colorjitter_factor: 0.5
-  colorjittor_p: 1
+  colorjitter_p: 1.O
-  # Range from which to sample colorjittor factor
+  sharpness_factor: 3.0
-  sharpness_factor: 3 
   # Should be more than 1, setting parameter to 1 does not change the image
   sharpness_p: 0.5 
   blur_factor: 0.2
@@ -71,4 +70,4 @@ image_transform:
   blur_p: 0.5 
   n_subset: 3
   # Maximum number of transforms to apply
-  list: ["colorjittor", "sharpness", "blur"]
+  list: ["colorjitter", "sharpness", "blur"]

lerobot/scripts/show_image_transforms.py

@@ -0,0 +1,65 @@
+
+from lerobot.common.utils.utils import init_hydra_config
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.transforms import make_transforms
+
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+
+
+DEFAULT_CONFIG_PATH = "configs/default.yaml"
+
+def show_image_transforms(cfg, repo_id, episode_index=0, output_dir="outputs/show_image_transforms"):
+    """
+    Apply a series of image transformations to a frame from a dataset and save the transformed images.
+
+    Args:
+        cfg (ConfigNode): The configuration object containing the image transformation settings and the dataset to sample.
+        repo_id (str): The ID of the repository.
+        episode_index (int, optional): The index of the episode to use. Defaults to 0.
+        output_dir (str, optional): The directory to save the transformed images. Defaults to "outputs/show_image_transforms".
+    """
+  
+    dataset = LeRobotDataset(repo_id)
+
+    print(f"Getting frame from camera: {dataset.camera_keys[0]}")
+
+    # Get first frame of given episode
+    from_idx = dataset.episode_data_index["from"][episode_index].item()
+    frame = dataset[from_idx][dataset.camera_keys[0]]
+
+    Path(output_dir).mkdir(parents=True, exist_ok=True)
+    base_filename = f"{output_dir}/episode_{episode_index}"
+
+    # Apply each transformation and save the result
+    for transform in cfg.list:
+        cfg = init_hydra_config(
+            DEFAULT_CONFIG_PATH,
+            overrides=[
+                f"image_transform.list=[{transform}]",
+                "image_transform.enable=True",
+                "image_transform.n_subset=1",
+                f"image_transform.{transform}_p=1",
+            ])
+        
+        cfg = cfg.image_transform
+
+        t = make_transforms(cfg)
+        
+        # Apply transformation to frame
+        transformed_frame = t(frame)
+        transformed_frame = transformed_frame.permute(1, 2, 0).numpy()
+
+        # Save transformed frame
+        plt.imshow(transformed_frame)
+        plt.savefig(f'{base_filename}_max_transform_{transform}.png')
+        plt.close()
+
+    frame = frame.permute(1, 2, 0).numpy()
+    # Save original frame
+    plt.imshow(frame)
+    plt.savefig(f'{base_filename}_original.png')
+    plt.close()
+
+    print(f"Saved transformed images.")