Fix task creation with gt job and gt job frame access

changelog.d/20241004_144626_mzhiltso_fix_task_creation_with_gt_job.md

@@ -0,0 +1,6 @@
+### Fixed
+
+- Invalid chunks for GT jobs when `?number` is used in the request and task frame step > 1
+  (<https://github.com/cvat-ai/cvat/pull/8510>)
+- Invalid output of frames for specific GT frame requests with `api/jobs/{id}/data/?type=frame`
+  (<https://github.com/cvat-ai/cvat/pull/8510>)

cvat/apps/engine/cache.py

@@ -504,8 +504,7 @@ def get_frames():
                     frame_range = (
                         (
                             db_data.start_frame
-                            + chunk_number * db_data.chunk_size
-                            + chunk_frame_idx * frame_step
+                            + (chunk_number * db_data.chunk_size + chunk_frame_idx) * frame_step
                         )
                         for chunk_frame_idx in range(db_data.chunk_size)
                     )

cvat/apps/engine/frame_provider.py

@@ -485,7 +485,7 @@ def __len__(self):
         return self._db_segment.frame_count
 
     def validate_frame_number(self, frame_number: int) -> Tuple[int, int, int]:
-        frame_sequence = list(self._db_segment.frame_set)
+        frame_sequence = sorted(self._db_segment.frame_set)
         abs_frame_number = self._get_abs_frame_number(self._db_segment.task.data, frame_number)
         if abs_frame_number not in frame_sequence:
             raise ValidationError(f"Incorrect requested frame number: {frame_number}")

cvat/apps/engine/serializers.py

@@ -822,6 +822,7 @@ def create(self, validated_data):
                     "The number of validation frames is not specified"
                 )
 
+            task_frame_provider = TaskFrameProvider(task)
             seed = validated_data.pop("random_seed", None)
 
             # The RNG backend must not change to yield reproducible results,
@@ -832,7 +833,7 @@ def create(self, validated_data):
             frames: list[int] = []
             overlap = task.overlap
             for segment in task.segment_set.all():
-                segment_frames = set(segment.frame_set)
+                segment_frames = set(map(task_frame_provider.get_rel_frame_number, segment.frame_set))
                 selected_frames = segment_frames.intersection(frames)
                 selected_count = len(selected_frames)
 
@@ -841,12 +842,14 @@ def create(self, validated_data):
                     continue
 
                 selectable_segment_frames = set(
-                    sorted(segment.frame_set)[overlap * (segment.start_frame != 0) : ]
+                    sorted(segment_frames)[overlap * (segment.start_frame != 0) : ]
                 ).difference(selected_frames)
 
                 frames.extend(rng.choice(
                     tuple(selectable_segment_frames), size=missing_count, replace=False
                 ).tolist())
+
+            frames = list(map(task_frame_provider.get_abs_frame_number, frames))
         elif frame_selection_method == models.JobFrameSelectionMethod.MANUAL:
             frames = validated_data.pop("frames")
 

cvat/apps/engine/task.py

@@ -1398,6 +1398,9 @@ def _update_status(msg: str) -> None:
         seed = validation_params.get("random_seed")
         rng = random.Generator(random.MT19937(seed=seed))
 
+        def _to_rel_frame(abs_frame: int) -> int:
+            return (abs_frame - db_data.start_frame) // db_data.get_frame_step()
+
         match validation_params["frame_selection_method"]:
             case models.JobFrameSelectionMethod.RANDOM_UNIFORM:
                 all_frames = range(db_data.size)
@@ -1431,7 +1434,7 @@ def _update_status(msg: str) -> None:
                 validation_frames: list[int] = []
                 overlap = db_task.overlap
                 for segment in db_task.segment_set.all():
-                    segment_frames = set(segment.frame_set)
+                    segment_frames = set(map(_to_rel_frame, segment.frame_set))
                     selected_frames = segment_frames.intersection(validation_frames)
                     selected_count = len(selected_frames)
 
@@ -1440,7 +1443,7 @@ def _update_status(msg: str) -> None:
                         continue
 
                     selectable_segment_frames = set(
-                        sorted(segment.frame_set)[overlap * (segment.start_frame != 0) : ]
+                        sorted(segment_frames)[overlap * (segment.start_frame != 0) : ]
                     ).difference(selected_frames)
 
                     validation_frames.extend(rng.choice(
@@ -1457,7 +1460,7 @@ def _update_status(msg: str) -> None:
                     )
 
                 validation_frames: list[int] = []
-                known_frame_names = {frame.path: frame.frame for frame in images}
+                known_frame_names = {frame.path: _to_rel_frame(frame.frame) for frame in images}
                 unknown_requested_frames = []
                 for frame_filename in validation_params['frames']:
                     frame_id = known_frame_names.get(frame_filename)
@@ -1484,11 +1487,14 @@ def _update_status(msg: str) -> None:
     # TODO: refactor
     if hasattr(db_data, 'validation_layout'):
         if db_data.validation_layout.mode == models.ValidationMode.GT:
+            def _to_abs_frame(rel_frame: int) -> int:
+                return rel_frame * db_data.get_frame_step() + db_data.start_frame
+
             db_gt_segment = models.Segment(
                 task=db_task,
                 start_frame=0,
                 stop_frame=db_data.size - 1,
-                frames=db_data.validation_layout.frames,
+                frames=list(map(_to_abs_frame, db_data.validation_layout.frames)),
                 type=models.SegmentType.SPECIFIC_FRAMES,
             )
         elif db_data.validation_layout.mode == models.ValidationMode.GT_POOL:
@@ -1653,7 +1659,7 @@ def _get_frame_size(frame_tuple: Tuple[av.VideoFrame, Any, Any]) -> int:
                             (abs_frame_id - media_extractor.start) // media_extractor.step
                             for abs_frame_id in (
                                 frame_map.get(frame, frame)
-                                for frame in db_segment.frame_set
+                                for frame in sorted(db_segment.frame_set)
                             )
                         ),
                         lambda _: next(frame_counter) // db_data.chunk_size

tests/python/rest_api/test_analytics.py

@@ -59,7 +59,7 @@ def _create_project(user, spec, **kwargs):
         return project.id, response.headers.get("X-Request-Id")
 
     @pytest.fixture(autouse=True)
-    def setup(self, restore_clickhouse_db_per_function):
+    def setup(self, restore_clickhouse_db_per_function, restore_redis_inmem_per_function):
         project_spec = {
             "name": f"Test project created by {self._USERNAME}",
             "labels": [