Fix to_tensorflow bugs

This patch fixes 2 to_tensorflow bugs 1. to_tensorflow fails when the meta information for a tensor includes dtype="object" 2. to_tensorflow also fails when it gets shape=(1,) in meta and the actual object has multiple dimensions Signed-off-by: Aditya Srivastava <adityasrivastava301199@gmail.com>
Gyanachand1 · Oct 26, 2020 · b56a9a3 · b56a9a3
1 parent 77f54b4
commit b56a9a3
Show file tree

Hide file tree

Showing 2 changed files with 17 additions and 16 deletions.
diff --git a/examples/coco/upload_coco2017.py b/examples/coco/upload_coco2017.py
@@ -20,14 +20,14 @@ def __init__(self, args, tag):
 
     def meta(self):
         return {
-            "segmentation": {"shape": (1,), "dtype": "object", "chunksize": 1000},
-            "area": {"shape": (1,), "dtype": "object", "chunksize": 1000},
-            "iscrowd": {"shape": (1,), "dtype": "object", "chunksize": 1000},
+            "segmentation": {"shape": (1,), "dtype": "uint32", "chunksize": 1000},
+            "area": {"shape": (1,), "dtype": "uint32", "chunksize": 1000},
+            "iscrowd": {"shape": (1,), "dtype": "uint8", "chunksize": 1000},
             "image_id": {"shape": (1,), "dtype": "int64"},
-            "bbox": {"shape": (1,), "dtype": "object", "chunksize": 1000},
-            "category_id": {"shape": (1,), "dtype": "object", "chunksize": 1000},
-            "id": {"shape": (1,), "dtype": "object", "chunksize": 1000},
-            "image": {"shape": (1,), "dtype": "object", "chunksize": 100},
+            "bbox": {"shape": (1,), "dtype": "uint16", "chunksize": 1000},
+            "category_id": {"shape": (1,), "dtype": "uint32", "chunksize": 1000},
+            "id": {"shape": (1,), "dtype": "uint32", "chunksize": 1000},
+            "image": {"shape": (1,), "dtype": "uint32", "chunksize": 100},
         }
 
     def __call__(self, input):
@@ -36,21 +36,22 @@ def __call__(self, input):
             # print(f"Image id: {input['image_id']}")
             ds["image_id"] = input["image_id"]
             info = input["info"]
-            ds["image"] = np.empty(1, object)
+            ds["image"] = np.empty(1, "uint32")
             ds["image"][0] = np.array(
                 Image.open(
                     os.path.join(
                         self._args.dataset_path,
                         get_image_name(self._args, self._tag, input["image_id"]),
                     )
-                )
+                ),
+                dtype="uint32",
             )
-            ds["segmentation"] = np.empty(1, object)
-            ds["area"] = np.empty(1, object)
-            ds["iscrowd"] = np.empty(1, object)
-            ds["bbox"] = np.empty(1, object)
-            ds["category_id"] = np.empty(1, object)
-            ds["id"] = np.empty(1, object)
+            ds["segmentation"] = np.empty(1, "uint32")
+            ds["area"] = np.empty(1, "uint32")
+            ds["iscrowd"] = np.empty(1, "uint8")
+            ds["bbox"] = np.empty(1, "uint16")
+            ds["category_id"] = np.empty(1, "uint32")
+            ds["id"] = np.empty(1, "uint32")
 
             ds["segmentation"][0] = [anno["segmentation"] for anno in info]
             ds["area"][0] = [anno["area"] for anno in info]

diff --git a/hub/collections/dataset/core.py b/hub/collections/dataset/core.py
@@ -627,7 +627,7 @@ def tf_dtype(np_dtype):
         return tf.data.Dataset.from_generator(
             tf_gen,
             output_types=output_types,
-            output_shapes=output_shapes,
+            # output_shapes=output_shapes,
         )