Job previews (#112)

CHANGELOG.md

@@ -17,6 +17,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Support source and target storages (server part) (<https://github.com/cvat-ai/cvat/pull/28>)
 - Tests for import/export annotation, dataset, backup from/to cloud storage (<https://github.com/cvat-ai/cvat/pull/56>)
 - Added Python SDK package (`cvat-sdk`) (<https://github.com/cvat-ai/cvat/pull/7>)
+- Previews for jobs (<https://github.com/cvat-ai/cvat/pull/112>)
 
 ### Changed
 - Bumped nuclio version to 1.8.14 (<https://github.com/cvat-ai/cvat/pull/29>)

cvat/apps/engine/media_extractors.py

@@ -111,7 +111,7 @@ class IMediaReader(ABC):
         pass
 
     @abstractmethod
-    def get_preview(self):
+    def get_preview(self, frame):
         pass
 
     @abstractmethod
@@ -198,11 +198,11 @@ class ImageListReader(IMediaReader):
     def get_progress(self, pos):
         return (pos - self._start + 1) / (self._stop - self._start)
 
-    def get_preview(self):
+    def get_preview(self, frame):
         if self._dimension == DimensionType.DIM_3D:
             fp = open(os.path.join(os.path.dirname(__file__), 'assets/3d_preview.jpeg'), "rb")
         else:
-            fp = open(self._source_path[0], "rb")
+            fp = open(self._source_path[frame], "rb")
         return self._get_preview(fp)
 
     def get_image_size(self, i):
@@ -338,12 +338,13 @@ class ZipReader(ImageListReader):
     def __del__(self):
         self._zip_source.close()
 
-    def get_preview(self):
+    def get_preview(self, frame):
         if self._dimension == DimensionType.DIM_3D:
             # TODO
             fp = open(os.path.join(os.path.dirname(__file__), 'assets/3d_preview.jpeg'), "rb")
             return self._get_preview(fp)
-        io_image = io.BytesIO(self._zip_source.read(self._source_path[0]))
+
+        io_image = io.BytesIO(self._zip_source.read(self._source_path[frame]))
         return self._get_preview(io_image)
 
     def get_image_size(self, i):
@@ -453,19 +454,23 @@ class VideoReader(IMediaReader):
                 duration = duration_sec * tb_denominator
         return duration
 
-    def get_preview(self):
+    def get_preview(self, frame):
         container = self._get_av_container()
         stream = container.streams.video[0]
-        preview = next(container.decode(stream))
+        tb_denominator = stream.time_base.denominator
-        return self._get_preview(preview.to_image() if not stream.metadata.get('rotate') \
+        needed_time = int((frame / stream.guessed_rate) * tb_denominator)
-            else av.VideoFrame().from_ndarray(
+        container.seek(offset=needed_time, stream=stream)
-                rotate_image(
+        for packet in container.demux(stream):
-                    preview.to_ndarray(format='bgr24'),
+            for frame in packet.decode():
-                    360 - int(container.streams.video[0].metadata.get('rotate'))
+                return self._get_preview(frame.to_image() if not stream.metadata.get('rotate') \
-                ),
+                    else av.VideoFrame().from_ndarray(
-                format ='bgr24'
+                        rotate_image(
-            ).to_image()
+                            frame.to_ndarray(format='bgr24'),
-        )
+                            360 - int(container.streams.video[0].metadata.get('rotate'))
+                        ),
+                        format ='bgr24'
+                    ).to_image()
+                )
 
     def get_image_size(self, i):
         image = (next(iter(self)))[0]

cvat/apps/engine/migrations/0056_jobs_previews.py

@@ -0,0 +1,64 @@
+# Generated by Django 3.2.14 on 2022-07-18 06:37
+
+import os
+import shutil
+from django.db import migrations
+from django.conf import settings
+from cvat.apps.engine.log import get_logger
+
+MIGRATION_NAME = os.path.splitext(os.path.basename(__file__))[0]
+MIGRATION_LOG = os.path.join(settings.MIGRATIONS_LOGS_ROOT, f"{MIGRATION_NAME}.log")
+
+def _get_query_set(apps):
+    Job = apps.get_model("engine", "Job")
+    query_set = Job.objects.all()
+    return query_set
+
+def _get_job_preview_path(jid):
+    return os.path.join(settings.JOBS_ROOT, str(jid), "preview.jpeg")
+
+def _get_data_preview_path(did):
+    return os.path.join(settings.MEDIA_DATA_ROOT, str(did), "preview.jpeg")
+
+def create_previews(apps, schema_editor):
+    logger = get_logger(MIGRATION_NAME, MIGRATION_LOG)
+    query_set = _get_query_set(apps)
+    logger.info(f'Migration has been started. Need to create {query_set.count()} previews.')
+    for db_job in query_set:
+        try:
+            jid = db_job.id
+            did = db_job.segment.task.data.id
+            task_preview = _get_data_preview_path(did)
+            job_preview =  _get_job_preview_path(jid)
+            if os.path.isfile(task_preview) and not os.path.isfile(job_preview):
+                shutil.copy(task_preview, job_preview)
+        except Exception as e:
+            logger.error(f'Cannot create preview for job {db_job.id}')
+            logger.error(str(e))
+
+def delete_previews(apps, schema_editor):
+    logger = get_logger(MIGRATION_NAME, MIGRATION_LOG)
+    query_set = _get_query_set(apps)
+    logger.info(f'Reverse migration has been started. Need to delete {query_set.count()} previews.')
+    for db_job in query_set:
+        try:
+            jid = db_job.id
+            job_preview =  _get_job_preview_path(jid)
+            if os.path.isfile(job_preview):
+                os.remove(job_preview)
+        except Exception as e:
+            logger.error(f'Cannot delete preview for job {db_job.id}')
+            logger.error(str(e))
+
+class Migration(migrations.Migration):
+
+    dependencies = [
+        ('engine', '0055_jobs_directories'),
+    ]
+
+    operations = [
+        migrations.RunPython(
+            code=create_previews,
+            reverse_code=delete_previews
+        )
+    ]

cvat/apps/engine/models.py

@@ -463,6 +463,8 @@ class Job(models.Model):
             })
         db_commit.save()
 
+    def get_preview_path(self):
+        return os.path.join(self.get_dirname(), "preview.jpeg")
 
     class Meta:
         default_permissions = ()

cvat/apps/engine/task.py

@@ -75,27 +75,31 @@ def _copy_data_from_source(server_files, upload_dir, server_dir=None):
                 os.makedirs(target_dir)
             shutil.copyfile(source_path, target_path)
 
-def _save_task_to_db(db_task):
+def _get_task_segment_data(db_task, data_size):
-    job = rq.get_current_job()
-    job.meta['status'] = 'Task is being saved in database'
-    job.save_meta()
-
     segment_size = db_task.segment_size
     segment_step = segment_size
-    if segment_size == 0 or segment_size > db_task.data.size:
+    if segment_size == 0 or segment_size > data_size:
-        segment_size = db_task.data.size
+        segment_size = data_size
-        db_task.segment_size = segment_size
 
         # Segment step must be more than segment_size + overlap in single-segment tasks
         # Otherwise a task contains an extra segment
         segment_step = sys.maxsize
 
-    default_overlap = 5 if db_task.mode == 'interpolation' else 0
+    overlap = 5 if db_task.mode == 'interpolation' else 0
-    if db_task.overlap is None:
+    if db_task.overlap is not None:
-        db_task.overlap = default_overlap
+        overlap = min(db_task.overlap, segment_size  // 2)
-    db_task.overlap = min(db_task.overlap, segment_size  // 2)
+
+    segment_step -= overlap
+    return segment_step, segment_size, overlap
 
-    segment_step -= db_task.overlap
+def _save_task_to_db(db_task, extractor):
+    job = rq.get_current_job()
+    job.meta['status'] = 'Task is being saved in database'
+    job.save_meta()
+
+    segment_step, segment_size, overlap = _get_task_segment_data(db_task, db_task.data.size)
+    db_task.segment_size = segment_size
+    db_task.overlap = overlap
 
     for start_frame in range(0, db_task.data.size, segment_step):
         stop_frame = min(start_frame + segment_size - 1, db_task.data.size - 1)
@@ -112,12 +116,13 @@ def _save_task_to_db(db_task):
         db_job = models.Job(segment=db_segment)
         db_job.save()
 
-        # create job directory
         job_path = db_job.get_dirname()
         if os.path.isdir(job_path):
             shutil.rmtree(job_path)
         os.makedirs(job_path)
 
+        preview = extractor.get_preview(frame=start_frame)
+        preview.save(db_job.get_preview_path())
 
     db_task.data.save()
     db_task.save()
@@ -328,8 +333,12 @@ def _create_thread(db_task, data, isBackupRestore=False, isDatasetImport=False):
         }
         cloud_storage_instance = get_cloud_storage_instance(cloud_provider=db_cloud_storage.provider_type, **details)
         sorted_media = sort(media['image'], data['sorting_method'])
-        first_sorted_media_image = sorted_media[0]
+
-        cloud_storage_instance.download_file(first_sorted_media_image, os.path.join(upload_dir, first_sorted_media_image))
+        data_size = len(sorted_media)
+        segment_step, *_ = _get_task_segment_data(db_task, data_size)
+        for start_frame in range(0, data_size, segment_step):
+            first_sorted_media_image = sorted_media[start_frame]
+            cloud_storage_instance.download_file(first_sorted_media_image, os.path.join(upload_dir, first_sorted_media_image))
 
         # prepare task manifest file from cloud storage manifest file
         # NOTE we should create manifest before defining chunk_size
@@ -665,8 +674,8 @@ def _create_thread(db_task, data, isBackupRestore=False, isDatasetImport=False):
         db_data.stop_frame = min(db_data.stop_frame, \
             db_data.start_frame + (db_data.size - 1) * db_data.get_frame_step())
 
-    preview = extractor.get_preview()
+    task_preview = extractor.get_preview(frame=0)
-    preview.save(db_data.get_preview_path())
+    task_preview.save(db_data.get_preview_path())
 
     slogger.glob.info("Found frames {} for Data #{}".format(db_data.size, db_data.id))
-    _save_task_to_db(db_task)
+    _save_task_to_db(db_task, extractor)

cvat/apps/engine/views.py

@@ -597,7 +597,7 @@ class DataChunkGetter:
         self.dimension = task_dim
 
 
-    def __call__(self, request, start, stop, db_data):
+    def __call__(self, request, start, stop, db_data, db_object):
         if not db_data:
             raise NotFound(detail='Cannot find requested data')
 
@@ -630,7 +630,7 @@ class DataChunkGetter:
             return HttpResponse(buf.getvalue(), content_type=mime)
 
         elif self.type == 'preview':
-            return sendfile(request, frame_provider.get_preview())
+            return sendfile(request, db_object.get_preview_path())
 
         elif self.type == 'context_image':
             if not (start <= self.number <= stop):
@@ -934,7 +934,7 @@ class TaskViewSet(UploadMixin, AnnotationMixin, viewsets.ModelViewSet, Serialize
                 self._object.dimension)
 
             return data_getter(request, self._object.data.start_frame,
-                self._object.data.stop_frame, self._object.data)
+                self._object.data.stop_frame, self._object.data, self._object.data)
 
     @extend_schema(methods=['PATCH'],
         operation_id='tasks_partial_update_data_file',
@@ -1492,7 +1492,7 @@ class JobViewSet(viewsets.GenericViewSet, mixins.ListModelMixin,
             db_job.segment.task.dimension)
 
         return data_getter(request, db_job.segment.start_frame,
-            db_job.segment.stop_frame, db_job.segment.task.data)
+            db_job.segment.stop_frame, db_job.segment.task.data, db_job)
 
     @extend_schema(summary='Method provides a meta information about media files which are related with the job',
         responses={
@@ -1988,7 +1988,7 @@ class CloudStorageViewSet(viewsets.ModelViewSet):
                 with NamedTemporaryFile() as temp_image:
                     storage.download_file(preview_path, temp_image.name)
                     reader = ImageListReader([temp_image.name])
-                    preview = reader.get_preview()
+                    preview = reader.get_preview(frame=0)
                     preview.save(db_storage.get_preview_path())
             content_type = mimetypes.guess_type(db_storage.get_preview_path())[0]
             return HttpResponse(open(db_storage.get_preview_path(), 'rb').read(), content_type)