Merge pull request #1864 from voxel51/feature/label-dict-imports

Adding support for label field dicts when importing labeled datasets

docs/source/user_guide/dataset_creation/datasets.rst

@@ -5164,14 +5164,20 @@ should implement is determined by the type of dataset that you are importing.
         importer = CustomLabeledImageDatasetImporter(...)
         label_field = ...
 
+        if isinstance(label_field, dict):
+            label_key = lambda k: label_field.get(k, k)
+        elif label_field is not None:
+            label_key = lambda k: label_field + "_" + k
+        else:
+            label_field = "ground_truth"
+            label_key = lambda k: k
+
         with importer:
             for image_path, image_metadata, label in importer:
                 sample = fo.Sample(filepath=image_path, metadata=image_metadata)
 
                 if isinstance(label, dict):
-                    sample.update_fields(
-                        {label_field + "_" + k: v for k, v in label.items()}
-                    )
+                    sample.update_fields({label_key(k): v for k, v in label.items()})
                 elif label is not None:
                     sample[label_field] = label
 
@@ -5566,14 +5572,20 @@ should implement is determined by the type of dataset that you are importing.
         importer = CustomLabeledVideoDatasetImporter(...)
         label_field = ...
 
+        if isinstance(label_field, dict):
+            label_key = lambda k: label_field.get(k, k)
+        elif label_field is not None:
+            label_key = lambda k: label_field + "_" + k
+        else:
+            label_field = "ground_truth"
+            label_key = lambda k: k
+
         with importer:
             for video_path, video_metadata, label, frames in importer:
                 sample = fo.Sample(filepath=video_path, metadata=video_metadata)
 
                 if isinstance(label, dict):
-                    sample.update_fields(
-                        {label_field + "_" + k: v for k, v in label.items()}
-                    )
+                    sample.update_fields({label_key(k): v for k, v in label.items()})
                 elif label is not None:
                     sample[label_field] = label
 
@@ -5583,8 +5595,7 @@ should implement is determined by the type of dataset that you are importing.
                     for frame_number, _label in frames.items():
                         if isinstance(_label, dict):
                             frame_labels[frame_number] = {
-                                label_field + "_" + field_name: label
-                                for field_name, label in _label.items()
+                                label_key(k): v for k, v in _label.items()
                             }
                         elif _label is not None:
                             frame_labels[frame_number] = {label_field: _label}

docs/source/user_guide/dataset_creation/samples.rst

@@ -932,6 +932,14 @@ classification or object detections) associated with the image.
             sample_parser = CustomLabeledImageSampleParser(...)
             label_field = ...
 
+            if isinstance(label_field, dict):
+                label_key = lambda k: label_field.get(k, k)
+            elif label_field is not None:
+                label_key = lambda k: label_field + "_" + k
+            else:
+                label_field = "ground_truth"
+                label_key = lambda k: k
+
             for sample in samples:
                 sample_parser.with_sample(sample)
 
@@ -947,9 +955,7 @@ classification or object detections) associated with the image.
                 sample = fo.Sample(filepath=image_path, metadata=metadata)
 
                 if isinstance(label, dict):
-                    sample.update_fields(
-                        {label_field + "_" + k: v for k, v in label.items()}
-                    )
+                    sample.update_fields({label_key(k): v for k, v in label.items()})
                 elif label is not None:
                     sample[label_field] = label
 
@@ -1191,6 +1197,14 @@ classification or object detections) associated with the image.
             sample_parser = CustomLabeledVideoSampleParser(...)
             label_field = ...
 
+            if isinstance(label_field, dict):
+                label_key = lambda k: label_field.get(k, k)
+            elif label_field is not None:
+                label_key = lambda k: label_field + "_" + k
+            else:
+                label_field = "ground_truth"
+                label_key = lambda k: k
+
             for sample in samples:
                 sample_parser.with_sample(sample)
 
@@ -1207,9 +1221,7 @@ classification or object detections) associated with the image.
                 sample = fo.Sample(filepath=video_path, metadata=metadata)
 
                 if isinstance(label, dict):
-                    sample.update_fields(
-                        {label_field + "_" + k: v for k, v in label.items()}
-                    )
+                    sample.update_fields({label_key(k): v for k, v in label.items()})
                 elif label is not None:
                     sample[label_field] = label
 
@@ -1219,8 +1231,7 @@ classification or object detections) associated with the image.
                     for frame_number, _label in frames.items():
                         if isinstance(_label, dict):
                             frame_labels[frame_number] = {
-                                label_field + "_" + field_name: label
-                                for field_name, label in _label.items()
+                                label_key(k): v for k, v in _label.items()
                             }
                         elif _label is not None:
                             frame_labels[frame_number] = {label_field: _label}

docs/source/user_guide/model_zoo/index.rst

@@ -225,7 +225,7 @@ All models in the FiftyOne Model Zoo are instances of the |Model| class, which
 defines a common interface for loading models and generating predictions with
 defined input and output data formats.
 
-.. note:
+.. note::
 
     The following sections describe the interface that all models in the Model
     Zoo implement. If you write a wrapper for your custom model that implements
@@ -357,7 +357,7 @@ and the output ``labels`` can be any of the following:
 
     # Multiple sample-level labels
     for key, value in labels.items():
-        sample[label_field + "_" + key] = value
+        sample[label_key(key)] = value
 
 -   A dict mapping frame numbers to |Label| instances. In this case, the
     provided labels are interpreted as frame-level labels that should be added
@@ -384,27 +384,37 @@ and the output ``labels`` can be any of the following:
     # Multiple per-frame labels
     sample.frames.merge(
         {
-            frame_number: {
-                label_field + "_" + name: label
-                for name, label in frame_dict.items()
-            }
+            frame_number: {label_key(k): v for k, v in frame_dict.items()}
             for frame_number, frame_dict in labels.items()
         }
     )
 
+In the above snippets, the ``label_key`` function maps label dict keys to field
+names, and is defined from ``label_field`` as follows::
+
+.. code-block:: python
+    :linenos:
+
+    if isinstance(label_field, dict):
+        label_key = lambda k: label_field.get(k, k)
+    elif label_field is not None:
+        label_key = lambda k: label_field + "_" + k
+    else:
+        label_key = lambda k: k
+
 For models that support batching, the |Model| interface also provides a
 :meth:`predict_all() <fiftyone.core.models.Model.predict_all>` method that can
 provide an efficient implementation of predicting on a batch of data.
 
-.. note:
+.. note::
 
     Builtin methods like
     :meth:`apply_model() <fiftyone.core.collections.SampleCollection.apply_model>`
     provide a ``batch_size`` parameter that can be used to control the batch
     size used when performing inference with models that support efficient
     batching.
 
-.. note:
+.. note::
 
     PyTorch models can implement the |TorchModelMixin| mixin, in which case
     `DataLoaders <https://pytorch.org/docs/stable/data.html#torch.utils.data.DataLoader>`_
@@ -440,7 +450,7 @@ By convention,
 :meth:`Model.embed() <fiftyone.core.models.EmbeddingsMixin.embed>` should
 return a numpy array containing the embedding.
 
-.. note:
+.. note::
 
     Sample embeddings are typically 1D vectors, but this is not strictly
     required.