Tensor shape and dimension names are now separate arrow fields

crates/store/re_chunk/src/iter.rs

@@ -710,7 +710,7 @@ impl Chunk {
             Err(err) => {
                 if cfg!(debug_assertions) {
                     panic!(
-                        "deserialization failed for {}, data discarded: {}",
+                        "[DEBUG-ONLY] deserialization failed for {}, data discarded: {}",
                         C::name(),
                         re_error::format_ref(&err),
                     );

crates/store/re_types/definitions/rerun/datatypes/tensor_data.fbs

@@ -19,8 +19,15 @@ table TensorData (
   "attr.python.array_aliases": "npt.ArrayLike",
   "attr.rust.derive": "PartialEq,"
 ) {
-  /// The shape of the tensor, including optional names for each dimension.
-  shape: [rerun.datatypes.TensorDimension] (order: 200);
+  /// The shape of the tensor, i.e. the length of each dimension.
+  shape: [uint64] (order: 200);
+
+  /// The names of the dimensions of the tensor (optional).
+  ///
+  /// If set, should be the same length as [datatypes.TensorData.shape].
+  ///
+  /// Example: `["height", "width", "channel", "batch"]`.
+  names: [string] (order: 250, nullable);
 
   /// The content/data.
   buffer: rerun.datatypes.TensorBuffer (order: 300);

crates/store/re_types/src/archetypes/depth_image_ext.rs

@@ -20,21 +20,19 @@ impl DepthImage {
         let tensor_data: TensorData = data
             .try_into()
             .map_err(ImageConstructionError::TensorDataConversion)?;
-        let shape = tensor_data.shape;
+        let TensorData { shape, buffer, .. } = tensor_data;
 
         let non_empty_dim_inds = find_non_empty_dim_indices(&shape);
 
         if non_empty_dim_inds.len() != 2 {
             return Err(ImageConstructionError::BadImageShape(shape));
         }
 
-        let (blob, datatype) = blob_and_datatype_from_tensor(tensor_data.buffer);
+        let (blob, datatype) = blob_and_datatype_from_tensor(buffer);
 
-        let (height, width) = (&shape[non_empty_dim_inds[0]], &shape[non_empty_dim_inds[1]]);
-        let height = height.size as u32;
-        let width = width.size as u32;
+        let (height, width) = (shape[non_empty_dim_inds[0]], shape[non_empty_dim_inds[1]]);
 
-        let image_format = ImageFormat::depth([width, height], datatype);
+        let image_format = ImageFormat::depth([width as u32, height as u32], datatype);
 
         Ok(Self {
             buffer: blob.into(),

crates/store/re_types/src/archetypes/image_ext.rs

@@ -33,33 +33,31 @@ impl Image {
         let tensor_data: TensorData = data
             .try_into()
             .map_err(ImageConstructionError::TensorDataConversion)?;
-        let shape = tensor_data.shape;
+        let TensorData { shape, buffer, .. } = tensor_data;
 
         let non_empty_dim_inds = find_non_empty_dim_indices(&shape);
 
         let is_shape_correct = match color_model {
             ColorModel::L => non_empty_dim_inds.len() == 2,
             ColorModel::RGB | ColorModel::BGR => {
-                non_empty_dim_inds.len() == 3 && shape[non_empty_dim_inds[2]].size == 3
+                non_empty_dim_inds.len() == 3 && shape[non_empty_dim_inds[2]] == 3
             }
             ColorModel::RGBA | ColorModel::BGRA => {
-                non_empty_dim_inds.len() == 3 && shape[non_empty_dim_inds[2]].size == 4
+                non_empty_dim_inds.len() == 3 && shape[non_empty_dim_inds[2]] == 4
             }
         };
 
         if !is_shape_correct {
             return Err(ImageConstructionError::BadImageShape(shape));
         }
 
-        let (blob, datatype) = blob_and_datatype_from_tensor(tensor_data.buffer);
+        let (blob, datatype) = blob_and_datatype_from_tensor(buffer);
 
-        let (height, width) = (&shape[non_empty_dim_inds[0]], &shape[non_empty_dim_inds[1]]);
-        let height = height.size as u32;
-        let width = width.size as u32;
+        let (height, width) = (shape[non_empty_dim_inds[0]], shape[non_empty_dim_inds[1]]);
 
         let image_format = ImageFormat {
-            width,
-            height,
+            width: width as _,
+            height: height as _,
             pixel_format: None,
             channel_datatype: Some(datatype),
             color_model: Some(color_model),

crates/store/re_types/src/archetypes/segmentation_image_ext.rs

@@ -19,21 +19,19 @@ impl SegmentationImage {
         let tensor_data: TensorData = data
             .try_into()
             .map_err(ImageConstructionError::TensorDataConversion)?;
-        let shape = tensor_data.shape;
+        let TensorData { shape, buffer, .. } = tensor_data;
 
         let non_empty_dim_inds = find_non_empty_dim_indices(&shape);
 
         if non_empty_dim_inds.len() != 2 {
             return Err(ImageConstructionError::BadImageShape(shape));
         }
 
-        let (blob, datatype) = blob_and_datatype_from_tensor(tensor_data.buffer);
+        let (blob, datatype) = blob_and_datatype_from_tensor(buffer);
 
-        let (height, width) = (&shape[non_empty_dim_inds[0]], &shape[non_empty_dim_inds[1]]);
-        let height = height.size as u32;
-        let width = width.size as u32;
+        let (height, width) = (shape[non_empty_dim_inds[0]], shape[non_empty_dim_inds[1]]);
 
-        let image_format = ImageFormat::segmentation([width, height], datatype);
+        let image_format = ImageFormat::segmentation([width as _, height as _], datatype);
 
         Ok(Self {
             buffer: blob.into(),

crates/store/re_types/src/archetypes/tensor_ext.rs

@@ -1,4 +1,4 @@
-use crate::datatypes::{TensorData, TensorDimension};
+use crate::datatypes::TensorData;
 
 use re_types_core::ArrowString;
 
@@ -24,37 +24,15 @@ impl Tensor {
 
     /// Update the `names` of the contained [`TensorData`] dimensions.
     ///
-    /// Any existing Dimension names will be overwritten.
+    /// Any existing names will be overwritten.
     ///
-    /// If too many, or too few names are provided, this function will warn and only
-    /// update the subset of names that it can.
-    pub fn with_dim_names(self, names: impl IntoIterator<Item = impl Into<ArrowString>>) -> Self {
-        let names: Vec<_> = names.into_iter().map(|x| Some(x.into())).collect();
-        if names.len() != self.data.0.shape.len() {
-            re_log::warn_once!(
-                "Wrong number of names provided for tensor dimension. {} provided but {} expected.",
-                names.len(),
-                self.data.0.shape.len(),
-            );
-        }
-        Self {
-            data: TensorData {
-                shape: self
-                    .data
-                    .0
-                    .shape
-                    .into_iter()
-                    .zip(names.into_iter().chain(std::iter::repeat(None)))
-                    .map(|(dim, name)| TensorDimension {
-                        size: dim.size,
-                        name: name.or(dim.name),
-                    })
-                    .collect(),
-                buffer: self.data.0.buffer,
-            }
-            .into(),
-            value_range: None,
-        }
+    /// If too few or too many names are provided, this function will warn and return.
+    pub fn with_dim_names(
+        mut self,
+        names: impl IntoIterator<Item = impl Into<ArrowString>>,
+    ) -> Self {
+        self.data.0 = self.data.0.with_dim_names(names);
+        self
     }
 }