diff --git a/src/encord_active/lib/embeddings/cnn.py b/src/encord_active/lib/embeddings/cnn.py
index 8bcc20d93..1225d67bd 100644
--- a/src/encord_active/lib/embeddings/cnn.py
+++ b/src/encord_active/lib/embeddings/cnn.py
@@ -88,7 +88,7 @@ def assemble_object_batch(data_unit: dict, img_path: Path, transforms: Optional[
 
 
 @torch.inference_mode()
-def generate_cnn_image_embeddings(iterator: Iterator, filepath: str) -> None:
+def generate_cnn_image_embeddings(iterator: Iterator) -> List[LabelEmbedding]:
     start = time.perf_counter()
     feature_extractor, transforms = get_model_and_transforms()
 
@@ -117,16 +117,15 @@ def generate_cnn_image_embeddings(iterator: Iterator, filepath: str) -> None:
         )
         collections.append(entry)
 
-    with open(filepath, "wb") as f:
-        pickle.dump(collections, f)
-
     logger.info(
         f"Generating {len(iterator)} embeddings took {str(time.perf_counter() - start)} seconds",
     )
 
+    return collections
+
 
 @torch.inference_mode()
-def generate_cnn_object_embeddings(iterator: Iterator, filepath: str) -> None:
+def generate_cnn_object_embeddings(iterator: Iterator) -> List[LabelEmbedding]:
     start = time.perf_counter()
     feature_extractor, transforms = get_model_and_transforms()
 
@@ -163,16 +162,15 @@ def generate_cnn_object_embeddings(iterator: Iterator, filepath: str) -> None:
 
                 collections.append(entry)
 
-    with open(filepath, "wb") as f:
-        pickle.dump(collections, f)
-
     logger.info(
         f"Generating {len(iterator)} embeddings took {str(time.perf_counter() - start)} seconds",
     )
 
+    return collections
+
 
 @torch.inference_mode()
-def generate_cnn_classification_embeddings(iterator: Iterator, filepath: str) -> None:
+def generate_cnn_classification_embeddings(iterator: Iterator) -> List[LabelEmbedding]:
     image_collections = get_cnn_embeddings(iterator, embedding_type=EmbeddingType.IMAGE)
 
     ontology_class_hash_to_index: dict[str, dict] = {}
@@ -259,13 +257,12 @@ def generate_cnn_classification_embeddings(iterator: Iterator, filepath: str) ->
             )
             collections.append(entry)
 
-    with open(filepath, "wb") as f:
-        pickle.dump(collections, f)
-
     logger.info(
         f"Generating {len(iterator)} embeddings took {str(time.perf_counter() - start)} seconds",
     )
 
+    return collections
+
 
 def get_cnn_embeddings(
     iterator: Iterator, embedding_type: EmbeddingType, *, force: bool = False
@@ -292,14 +289,18 @@ def get_cnn_embeddings(
 
 def generate_cnn_embeddings(iterator: Iterator, embedding_type: EmbeddingType, target: str):
     if embedding_type == EmbeddingType.IMAGE:
-        generate_cnn_image_embeddings(iterator, filepath=target)
+        cnn_embeddings = generate_cnn_image_embeddings(iterator)
     elif embedding_type == EmbeddingType.OBJECT:
-        generate_cnn_object_embeddings(iterator, filepath=target)
+        cnn_embeddings = generate_cnn_object_embeddings(iterator)
     elif embedding_type == EmbeddingType.CLASSIFICATION:
-        generate_cnn_classification_embeddings(iterator, filepath=target)
+        cnn_embeddings = generate_cnn_classification_embeddings(iterator)
+    else:
+        raise ValueError(f"Unsupported embedding type {embedding_type}")
+
+    target_path = Path(target)
+    target_path.parent.mkdir(parents=True, exist_ok=True)
+    target_path.write_bytes(pickle.dumps(cnn_embeddings))
 
-    with open(target, "rb") as f:
-        cnn_embeddings = pickle.load(f)
     logger.info("Done!")
 
     return cnn_embeddings