Drop support of the TFRecord format (#7416)

Usage statistics on app.cvat.ai show that this format is rarely used, with single-digit numbers of projects/tasks/jobs being exported or imported using this format. Moreover, TensorFlow's popularity appears to be shrinking, so I'm not expecting this format to make a comeback. Meanwhile, supporting this format has a cost that has to be borne by everyone deploying CVAT, because it requires TensorFlow to be installed and loaded. This has various ill effects: 1. Loading time is increased. In my testing, even a command as trivial as `manage.py --help` is slowed down by 3.6 seconds. This may not seem like much, but the effect is compounded, because we have multiple processes (server + workers) all loading the same codebase. Plus, the container entrypoint may execute several Django commands. 2. Memory usage is increased. TensorFlow adds ~100MB of RAM usage per process with data alone; and the libraries add more (although it's hard to estimate the impact of library code, since it can be shared between processes in RAM). 3. Docker image size is increased by ~1.5GB (when unpacked). This is more than half of the current total size! Building time is increased as well. Overall, it seems that the drawbacks of keeping support for this format outweigh the benefits, so it's time to drop it.
cvat-ai · Feb 19, 2024 · 7286d65 · 7286d65
1 parent f0af22a
commit 7286d65
Show file tree

Hide file tree

Showing 17 changed files with 18 additions and 378 deletions.
diff --git a/README.md b/README.md
@@ -155,7 +155,6 @@ For more information about the supported formats, see:
 | [YOLO](https://pjreddie.com/darknet/yolo/)                                                       | ✔️     | ✔️     |
 | [MS COCO Object Detection](http://cocodataset.org/#format-data)                                  | ✔️     | ✔️     |
 | [MS COCO Keypoints Detection](http://cocodataset.org/#format-data)                               | ✔️     | ✔️     |
-| [TFrecord](https://www.tensorflow.org/tutorials/load_data/tfrecord)                              | ✔️     | ✔️     |
 | [MOT](https://motchallenge.net/)                                                                 | ✔️     | ✔️     |
 | [MOTS PNG](https://www.vision.rwth-aachen.de/page/mots)                                          | ✔️     | ✔️     |
 | [LabelMe 3.0](http://labelme.csail.mit.edu/Release3.0)                                           | ✔️     | ✔️     |

diff --git a/changelog.d/20240131_180300_roman_remove_tfrecord.md b/changelog.d/20240131_180300_roman_remove_tfrecord.md
@@ -0,0 +1,4 @@
+### Removed
+
+- Removed support for the TFRecord dataset format
+  (<https://github.com/opencv/cvat/pull/7416>)
diff --git a/cvat/apps/dataset_manager/formats/registry.py b/cvat/apps/dataset_manager/formats/registry.py
@@ -95,13 +95,6 @@ def make_exporter(name):
     return EXPORT_FORMATS[name]()
 
 
-# Add checking for TF availability to avoid CVAT sever instance / interpreter
-# crash and provide a meaningful diagnostic message in the case of AVX
-# instructions unavailability:
-# https://github.com/openvinotoolkit/cvat/pull/1567
-import datumaro.util.tf_util as TF
-TF.enable_tf_check = True
-
 # pylint: disable=unused-import
 import cvat.apps.dataset_manager.formats.coco
 import cvat.apps.dataset_manager.formats.cvat
@@ -111,7 +104,6 @@ def make_exporter(name):
 import cvat.apps.dataset_manager.formats.mot
 import cvat.apps.dataset_manager.formats.mots
 import cvat.apps.dataset_manager.formats.pascal_voc
-import cvat.apps.dataset_manager.formats.tfrecord
 import cvat.apps.dataset_manager.formats.yolo
 import cvat.apps.dataset_manager.formats.imagenet
 import cvat.apps.dataset_manager.formats.camvid

diff --git a/cvat/apps/dataset_manager/formats/tfrecord.py b/cvat/apps/dataset_manager/formats/tfrecord.py
diff --git a/cvat/apps/dataset_manager/tests/assets/annotations.json b/cvat/apps/dataset_manager/tests/assets/annotations.json
@@ -701,24 +701,6 @@
     ],
     "tracks": []
   },
-  "TFRecord 1.0": {
-    "version": 0,
-    "tags": [],
-    "shapes": [
-      {
-        "type": "rectangle",
-        "occluded": false,
-        "z_order": 0,
-        "points": [16.5, 17.2, 38.89, 25.63],
-        "frame": 0,
-        "label_id": null,
-        "group": 0,
-        "source": "manual",
-        "attributes": []
-      }
-    ],
-    "tracks": []
-  },
   "YOLO 1.1": {
     "version": 0,
     "tags": [],

diff --git a/cvat/apps/dataset_manager/tests/test_formats.py b/cvat/apps/dataset_manager/tests/test_formats.py
@@ -295,7 +295,6 @@ def test_export_formats_query(self):
             'MOTS PNG 1.0',
             'PASCAL VOC 1.1',
             'Segmentation mask 1.1',
-            'TFRecord 1.0',
             'YOLO 1.1',
             'ImageNet 1.0',
             'CamVid 1.0',
@@ -326,7 +325,6 @@ def test_import_formats_query(self):
             'MOTS PNG 1.0',
             'PASCAL VOC 1.1',
             'Segmentation mask 1.1',
-            'TFRecord 1.0',
             'YOLO 1.1',
             'ImageNet 1.0',
             'CamVid 1.0',
@@ -381,7 +379,6 @@ def test_empty_images_are_exported(self):
             # ('MOTS PNG 1.0', 'mots_png'), # does not support
             ('PASCAL VOC 1.1', 'voc'),
             ('Segmentation mask 1.1', 'voc'),
-            ('TFRecord 1.0', 'tf_detection_api'),
             ('YOLO 1.1', 'yolo'),
             ('ImageNet 1.0', 'imagenet_txt'),
             ('CamVid 1.0', 'camvid'),

diff --git a/cvat/apps/dataset_manager/tests/test_rest_api_formats.py b/cvat/apps/dataset_manager/tests/test_rest_api_formats.py
@@ -417,7 +417,7 @@ def test_api_v2_dump_and_upload_annotations_with_objects_type_is_shape(self):
                         "Cityscapes 1.0", "Datumaro 1.0",
                         "ImageNet 1.0", "MOTS PNG 1.0",
                         "PASCAL VOC 1.1", "Segmentation mask 1.1",
-                        "TFRecord 1.0", "VGGFace2 1.0",
+                        "VGGFace2 1.0",
                         "WiderFace 1.0", "YOLO 1.1"
                     ]:
                         self._create_annotations(task, dump_format_name, "default")
@@ -522,7 +522,7 @@ def test_api_v2_dump_annotations_with_objects_type_is_track(self):
                     if dump_format_name in [
                             "Cityscapes 1.0", "ImageNet 1.0",
                             "MOTS PNG 1.0", "PASCAL VOC 1.1",
-                            "Segmentation mask 1.1", "TFRecord 1.0",
+                            "Segmentation mask 1.1",
                             "VGGFace2 1.0", "WiderFace 1.0", "YOLO 1.1"
                     ]:
                         self._create_annotations(task, dump_format_name, "default")
@@ -968,7 +968,7 @@ def test_api_v2_rewriting_annotations(self):
 
                     if dump_format_name in [
                         "MOT 1.1", "PASCAL VOC 1.1", "Segmentation mask 1.1",
-                        "TFRecord 1.0", "YOLO 1.1", "ImageNet 1.0",
+                        "YOLO 1.1", "ImageNet 1.0",
                         "WiderFace 1.0", "VGGFace2 1.0",
                         "Datumaro 1.0", "Open Images V6 1.0", "KITTI 1.0"
                     ]:
@@ -1084,7 +1084,7 @@ def test_api_v2_tasks_annotations_dump_and_upload_with_datumaro(self):
                     if dump_format_name in [
                         "MOT 1.1", "MOTS PNG 1.0",
                         "PASCAL VOC 1.1", "Segmentation mask 1.1",
-                        "TFRecord 1.0", "YOLO 1.1", "ImageNet 1.0",
+                        "YOLO 1.1", "ImageNet 1.0",
                         "WiderFace 1.0", "VGGFace2 1.0", "LFW 1.0",
                         "Open Images V6 1.0", "Datumaro 1.0", "KITTI 1.0"
                     ]:
@@ -1287,7 +1287,7 @@ def test_api_v2_export_import_dataset(self):
                 dump_format_name = dump_format.DISPLAY_NAME
                 if dump_format_name in [
                     'Cityscapes 1.0', 'LFW 1.0', 'Market-1501 1.0',
-                    'MOT 1.1', 'TFRecord 1.0'
+                    'MOT 1.1',
                 ]:
                     # TO-DO: fix bug for this formats
                     continue
@@ -1303,7 +1303,7 @@ def test_api_v2_export_import_dataset(self):
                 if dump_format_name in [
                     "Cityscapes 1.0", "Datumaro 1.0", "ImageNet 1.0",
                     "MOT 1.1", "MOTS PNG 1.0", "PASCAL VOC 1.1",
-                    "Segmentation mask 1.1", "TFRecord 1.0", "VGGFace2 1.0",
+                    "Segmentation mask 1.1", "VGGFace2 1.0",
                     "WiderFace 1.0", "YOLO 1.1"
                 ]:
                     self._create_annotations(task, dump_format_name, "default")
@@ -1344,7 +1344,7 @@ def test_api_v2_export_import_dataset(self):
                 upload_format_name = upload_format.DISPLAY_NAME
                 if upload_format_name in [
                     'Cityscapes 1.0', 'LFW 1.0', 'Market-1501 1.0',
-                    'MOT 1.1', 'TFRecord 1.0'
+                    'MOT 1.1',
                 ]:
                     # TO-DO: fix bug for this formats
                     continue

diff --git a/cvat/apps/engine/tests/test_rest_api.py b/cvat/apps/engine/tests/test_rest_api.py
@@ -6031,8 +6031,7 @@ def _get_initial_annotation(annotation_format):
                 annotations["shapes"] = rectangle_shapes_wo_attrs
                 annotations["tags"] = tags_wo_attrs
 
-            elif annotation_format == "YOLO 1.1" or \
-                 annotation_format == "TFRecord 1.0":
+            elif annotation_format == "YOLO 1.1":
                 annotations["shapes"] = rectangle_shapes_wo_attrs
 
             elif annotation_format == "COCO 1.0":
@@ -6404,8 +6403,6 @@ def etree_to_dict(t):
                 for json in jsons:
                     coco = coco_loader.COCO(json)
                     self.assertTrue(coco.getAnnIds())
-        elif format_name == "TFRecord 1.0":
-            self.assertTrue(zipfile.is_zipfile(content))
         elif format_name == "Segmentation mask 1.1":
             self.assertTrue(zipfile.is_zipfile(content))
 

diff --git a/cvat/requirements/base.in b/cvat/requirements/base.in
@@ -51,4 +51,3 @@ rq-scheduler==0.13.1
 rq==1.15.1
 rules>=3.3
 Shapely==1.7.1
-tensorflow==2.11.1 # Optional requirement of Datumaro. Use tensorflow-macos==2.8.0 for Mac M1