Merge remote-tracking branch 'origin/main' into user/aliberts/2024_09_25_reshape_dataset

examples/10_use_so100.md

@@ -135,7 +135,7 @@ You will need to move the follower arm to these positions sequentially:
 Make sure both arms are connected and run this script to launch manual calibration:
 ```bash
 python lerobot/scripts/control_robot.py calibrate \
-    --robot-path lerobot/configs/robot/moss.yaml \
+    --robot-path lerobot/configs/robot/so100.yaml \
     --robot-overrides '~cameras' --arms main_follower
 ```
 

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -67,6 +67,7 @@ class DiffusionConfig:
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
             The group sizes are set to be about 16 (to be precise, feature_dim // 16).
         spatial_softmax_num_keypoints: Number of keypoints for SpatialSoftmax.
+        use_separate_rgb_encoders_per_camera: Whether to use a separate RGB encoder for each camera view.
         down_dims: Feature dimension for each stage of temporal downsampling in the diffusion modeling Unet.
             You may provide a variable number of dimensions, therefore also controlling the degree of
             downsampling.
@@ -130,6 +131,7 @@ class DiffusionConfig:
     pretrained_backbone_weights: str | None = None
     use_group_norm: bool = True
     spatial_softmax_num_keypoints: int = 32
+    use_separate_rgb_encoder_per_camera: bool = False
     # Unet.
     down_dims: tuple[int, ...] = (512, 1024, 2048)
     kernel_size: int = 5

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -182,6 +182,11 @@ class DiffusionModel(nn.Module):
         self._use_env_state = False
         if num_images > 0:
             self._use_images = True
+            if self.config.use_separate_rgb_encoder_per_camera:
+                encoders = [DiffusionRgbEncoder(config) for _ in range(num_images)]
+                self.rgb_encoder = nn.ModuleList(encoders)
+                global_cond_dim += encoders[0].feature_dim * num_images
+            else:
                 self.rgb_encoder = DiffusionRgbEncoder(config)
                 global_cond_dim += self.rgb_encoder.feature_dim * num_images
         if "observation.environment_state" in config.input_shapes:
@@ -239,8 +244,24 @@ class DiffusionModel(nn.Module):
         """Encode image features and concatenate them all together along with the state vector."""
         batch_size, n_obs_steps = batch["observation.state"].shape[:2]
         global_cond_feats = [batch["observation.state"]]
-        # Extract image feature (first combine batch, sequence, and camera index dims).
+        # Extract image features.
         if self._use_images:
+            if self.config.use_separate_rgb_encoder_per_camera:
+                # Combine batch and sequence dims while rearranging to make the camera index dimension first.
+                images_per_camera = einops.rearrange(batch["observation.images"], "b s n ... -> n (b s) ...")
+                img_features_list = torch.cat(
+                    [
+                        encoder(images)
+                        for encoder, images in zip(self.rgb_encoder, images_per_camera, strict=True)
+                    ]
+                )
+                # Separate batch and sequence dims back out. The camera index dim gets absorbed into the
+                # feature dim (effectively concatenating the camera features).
+                img_features = einops.rearrange(
+                    img_features_list, "(n b s) ... -> b s (n ...)", b=batch_size, s=n_obs_steps
+                )
+            else:
+                # Combine batch, sequence, and "which camera" dims before passing to shared encoder.
                 img_features = self.rgb_encoder(
                     einops.rearrange(batch["observation.images"], "b s n ... -> (b s n) ...")
                 )