def gen_video_preview(elem, output_dir):
"""
Copy temporary image to specified output filepath.
:param elem: Data element to get the preview image for.
:type elem: smqtk.data_rep.DataElement
:param output_dir: Directory to save generated image to.
:type output_dir: str
"""
output_fp = os.path.join(output_dir,
"%s.gif" % elem.md5())
if not os.path.isfile(output_fp):
tmp_vid_fp = elem.write_temp()
interval = 0.5 # ~2fps gif
fm = video_utils.ffmpeg_extract_frame_map(
tmp_vid_fp, second_interval=interval
)
img_arrays = []
for frm_num in sorted(fm.keys()):
img_arrays.append(imageio.imread(fm[frm_num]))
imageio.mimwrite(output_fp, img_arrays, duration=interval)
elem.clean_temp()
return output_fp
def gen_video_preview(elem, output_dir):
"""
Copy temporary image to specified output filepath.
:param elem: Data element to get the preview image for.
:type elem: smqtk.representation.DataElement
:param output_dir: Directory to save generated image to.
:type output_dir: str
"""
output_fp = os.path.join(output_dir, "%s.gif" % elem.uuid())
if not os.path.isfile(output_fp):
tmp_vid_fp = elem.write_temp()
interval = 0.5 # ~2fps gif
fm = video_utils.ffmpeg_extract_frame_map(tmp_vid_fp,
second_interval=interval)
img_arrays = []
for frm_num in sorted(fm.keys()):
img_arrays.append(imageio.imread(fm[frm_num]))
imageio.mimwrite(output_fp, img_arrays, duration=interval)
elem.clean_temp()
return output_fp
def _generate_descriptor_matrices(self, data_set, **kwargs):
"""
Generate info and descriptor matrices based on ingest type.
:param data_set: Iterable of data elements to generate combined info
and descriptor matrices for.
:type item_iter: collections.Set[smqtk.representation.DataElement]
:param limit: Limit the number of descriptor entries to this amount.
:type limit: int
:return: Combined info and descriptor matrices for all base images
:rtype: (numpy.core.multiarray.ndarray, numpy.core.multiarray.ndarray)
"""
descriptor_limit = kwargs.get('limit', float('inf'))
# With videos, an "item" is one video, so, collect for a while video
# as normal, then subsample from the full video collection.
per_item_limit = numpy.floor(float(descriptor_limit) / len(data_set))
# If an odd number of jobs, favor descriptor extraction
if self.parallel:
descr_parallel = int(max(1, math.ceil(self.parallel / 2.0)))
extract_parallel = int(max(1, math.floor(self.parallel / 2.0)))
else:
cpuc = multiprocessing.cpu_count()
descr_parallel = int(max(1, math.ceil(cpuc / 2.0)))
extract_parallel = int(max(1, math.floor(cpuc / 2.0)))
# For each video, extract frames and submit colorDescriptor processing
# jobs for each frame, combining all results into a single matrix for
# return.
pool = multiprocessing.Pool(processes=descr_parallel)
# Mapping of [UID] to [frame] to tuple containing:
# (info_fp, desc_fp, async processing result)
r_map = {}
with SimpleTimer("Extracting frames and submitting descriptor jobs...",
self._log.debug):
for di in data_set:
r_map[di.uuid()] = {}
tmp_vid_fp = self._get_data_temp_path(di)
p = dict(self.FRAME_EXTRACTION_PARAMS)
vmd = get_metadata_info(tmp_vid_fp)
p['second_offset'] = vmd.duration * p['second_offset']
p['max_duration'] = vmd.duration * p['max_duration']
fm = video_utils.ffmpeg_extract_frame_map(
self._work_dir, tmp_vid_fp, parallel=extract_parallel, **p)
# Compute descriptors for extracted frames.
for frame, imgPath in fm.iteritems():
info_fp, desc_fp = \
self._get_standard_info_descriptors_filepath(di, frame)
r = pool.apply_async(utils.generate_descriptors,
args=(self.EXE, imgPath,
self.descriptor_type(), info_fp,
desc_fp))
r_map[di.uuid()][frame] = (info_fp, desc_fp, r)
# Clean temporary video file file while computing descriptors
# This does not remove the extracted frames that the underlying
# detector/descriptor is working on.
di.clean_temp()
pool.close()
# Each result is a tuple of two ndarrays: info and descriptor matrices
with SimpleTimer("Collecting shape information for super matrices...",
self._log.debug):
running_height = 0
i_width = None
d_width = None
# Transform r_map[uid] into:
# (info_mat_files, desc_mat_files, sR, ssi_list)
# -> files in frame order
uids = sorted(r_map)
for uid in uids:
video_num_desc = 0
video_info_mat_fps = [] # ordered list of frame info mat files
video_desc_mat_fps = [] # ordered list of frame desc mat files
for frame in sorted(r_map[uid]):
ifp, dfp, r = r_map[uid][frame]
# Descriptor generation may have failed for this UID
try:
i_shape, d_shape = r.get()
except RuntimeError, ex:
self._log.warning(
'Descriptor generation failed for '
'frame %d in video UID[%s]: %s', frame, uid,
str(ex))
r_map[uid] = None
continue
if d_width is None and d_shape[0] != 0:
i_width = i_shape[1]
d_width = d_shape[1]
# Skip if there were no descriptors generated for this
# frame
if d_shape[1] == 0:
continue
video_info_mat_fps.append(ifp)
video_desc_mat_fps.append(dfp)
video_num_desc += d_shape[0]
# If combined descriptor height exceeds the per-item limit,
# generate a random subsample index list
ssi = None
if video_num_desc > per_item_limit:
ssi = sorted(
numpy.random.permutation(video_num_desc)
[:per_item_limit])
video_num_desc = len(ssi)
r_map[uid] = (video_info_mat_fps, video_desc_mat_fps,
running_height, ssi)
running_height += video_num_desc
def _generate_descriptor_matrices(self, data_set, **kwargs):
"""
Generate info and descriptor matrices based on ingest type.
:param data_set: Iterable of data elements to generate combined info
and descriptor matrices for.
:type item_iter: collections.Set[smqtk.data_rep.DataElement]
:param limit: Limit the number of descriptor entries to this amount.
:type limit: int
:return: Combined info and descriptor matrices for all base images
:rtype: (numpy.core.multiarray.ndarray, numpy.core.multiarray.ndarray)
"""
descriptor_limit = kwargs.get('limit', float('inf'))
# With videos, an "item" is one video, so, collect for a while video
# as normal, then subsample from the full video collection.
per_item_limit = numpy.floor(float(descriptor_limit) / len(data_set))
# If an odd number of jobs, favor descriptor extraction
if self.PARALLEL:
descr_parallel = int(max(1, math.ceil(self.PARALLEL/2.0)))
extract_parallel = int(max(1, math.floor(self.PARALLEL/2.0)))
else:
cpuc = multiprocessing.cpu_count()
descr_parallel = int(max(1, math.ceil(cpuc/2.0)))
extract_parallel = int(max(1, math.floor(cpuc/2.0)))
# For each video, extract frames and submit colorDescriptor processing
# jobs for each frame, combining all results into a single matrix for
# return.
pool = multiprocessing.Pool(processes=descr_parallel)
# Mapping of [UID] to [frame] to tuple containing:
# (info_fp, desc_fp, async processing result)
r_map = {}
with SimpleTimer("Extracting frames and submitting descriptor jobs...",
self.log.debug):
for di in data_set:
r_map[di.uuid()] = {}
tmp_vid_fp = di.write_temp(self.temp_dir)
p = dict(self.FRAME_EXTRACTION_PARAMS)
vmd = get_metadata_info(tmp_vid_fp)
p['second_offset'] = vmd.duration * p['second_offset']
p['max_duration'] = vmd.duration * p['max_duration']
fm = video_utils.ffmpeg_extract_frame_map(
tmp_vid_fp,
parallel=extract_parallel,
**p
)
# Compute descriptors for extracted frames.
for frame, imgPath in fm.iteritems():
info_fp, desc_fp = \
self._get_standard_info_descriptors_filepath(di, frame)
r = pool.apply_async(
utils.generate_descriptors,
args=(self.PROC_COLORDESCRIPTOR, imgPath,
self.descriptor_type(), info_fp, desc_fp)
)
r_map[di.uuid()][frame] = (info_fp, desc_fp, r)
# Clean temporary file while computing descriptors
di.clean_temp()
pool.close()
# Each result is a tuple of two ndarrays: info and descriptor matrices
with SimpleTimer("Collecting shape information for super matrices...",
self.log.debug):
running_height = 0
# Known constants
i_width = 5
d_width = 384
# Transform r_map[uid] into:
# (info_mat_files, desc_mat_files, sR, ssi_list)
# -> files in frame order
uids = sorted(r_map)
for uid in uids:
video_num_desc = 0
video_info_mat_fps = [] # ordered list of frame info mat files
video_desc_mat_fps = [] # ordered list of frame desc mat files
for frame in sorted(r_map[uid]):
ifp, dfp, r = r_map[uid][frame]
i_shape, d_shape = r.get()
if None in (i_width, d_width):
i_width = i_shape[1]
d_width = d_shape[1]
video_info_mat_fps.append(ifp)
video_desc_mat_fps.append(dfp)
video_num_desc += i_shape[0]
# If combined descriptor height exceeds the per-item limit,
# generate a random subsample index list
ssi = None
if video_num_desc > per_item_limit:
ssi = sorted(
numpy.random.permutation(video_num_desc)[:per_item_limit]
)
video_num_desc = len(ssi)
r_map[uid] = (video_info_mat_fps, video_desc_mat_fps,
running_height, ssi)
running_height += video_num_desc
pool.join()
del pool
with SimpleTimer("Building master descriptor matrices...",
self.log.debug):
master_info = numpy.zeros((running_height, i_width), dtype=float)
master_desc = numpy.zeros((running_height, d_width), dtype=float)
tp = multiprocessing.pool.ThreadPool(processes=self.PARALLEL)
for uid in uids:
info_fp_list, desc_fp_list, sR, ssi = r_map[uid]
tp.apply_async(ColorDescriptor_Video._thread_load_matrices,
args=(master_info, info_fp_list, sR, ssi))
tp.apply_async(ColorDescriptor_Video._thread_load_matrices,
args=(master_desc, desc_fp_list, sR, ssi))
tp.close()
tp.join()
return master_info, master_desc
def _generate_descriptor_matrices(self, data_set, **kwargs):
"""
Generate info and descriptor matrices based on ingest type.
:param data_set: Iterable of data elements to generate combined info
and descriptor matrices for.
:type item_iter: collections.Set[smqtk.representation.DataElement]
:param limit: Limit the number of descriptor entries to this amount.
:type limit: int
:return: Combined info and descriptor matrices for all base images
:rtype: (numpy.core.multiarray.ndarray, numpy.core.multiarray.ndarray)
"""
descriptor_limit = kwargs.get('limit', float('inf'))
# With videos, an "item" is one video, so, collect for a while video
# as normal, then subsample from the full video collection.
per_item_limit = numpy.floor(float(descriptor_limit) / len(data_set))
# If an odd number of jobs, favor descriptor extraction
if self.PARALLEL:
descr_parallel = int(max(1, math.ceil(self.PARALLEL/2.0)))
extract_parallel = int(max(1, math.floor(self.PARALLEL/2.0)))
else:
cpuc = multiprocessing.cpu_count()
descr_parallel = int(max(1, math.ceil(cpuc/2.0)))
extract_parallel = int(max(1, math.floor(cpuc/2.0)))
# For each video, extract frames and submit colorDescriptor processing
# jobs for each frame, combining all results into a single matrix for
# return.
pool = multiprocessing.Pool(processes=descr_parallel)
# Mapping of [UID] to [frame] to tuple containing:
# (info_fp, desc_fp, async processing result)
r_map = {}
with SimpleTimer("Extracting frames and submitting descriptor jobs...",
self._log.debug):
for di in data_set:
r_map[di.uuid()] = {}
tmp_vid_fp = self._get_data_temp_path(di)
p = dict(self.FRAME_EXTRACTION_PARAMS)
vmd = get_metadata_info(tmp_vid_fp)
p['second_offset'] = vmd.duration * p['second_offset']
p['max_duration'] = vmd.duration * p['max_duration']
fm = video_utils.ffmpeg_extract_frame_map(
self._work_dir,
tmp_vid_fp,
parallel=extract_parallel,
**p
)
# Compute descriptors for extracted frames.
for frame, imgPath in fm.iteritems():
info_fp, desc_fp = \
self._get_standard_info_descriptors_filepath(di, frame)
r = pool.apply_async(
utils.generate_descriptors,
args=(self.PROC_COLORDESCRIPTOR, imgPath,
self.descriptor_type(), info_fp, desc_fp)
)
r_map[di.uuid()][frame] = (info_fp, desc_fp, r)
# Clean temporary video file file while computing descriptors
# This does not remove the extracted frames that the underlying
# detector/descriptor is working on.
di.clean_temp()
pool.close()
# Each result is a tuple of two ndarrays: info and descriptor matrices
with SimpleTimer("Collecting shape information for super matrices...",
self._log.debug):
running_height = 0
i_width = None
d_width = None
# Transform r_map[uid] into:
# (info_mat_files, desc_mat_files, sR, ssi_list)
# -> files in frame order
uids = sorted(r_map)
for uid in uids:
video_num_desc = 0
video_info_mat_fps = [] # ordered list of frame info mat files
video_desc_mat_fps = [] # ordered list of frame desc mat files
for frame in sorted(r_map[uid]):
ifp, dfp, r = r_map[uid][frame]
# Descriptor generation may have failed for this UID
try:
i_shape, d_shape = r.get()
except RuntimeError, ex:
self._log.warning('Descriptor generation failed for '
'frame %d in video UID[%s]: %s',
frame, uid, str(ex))
r_map[uid] = None
continue
if d_width is None and d_shape[0] != 0:
i_width = i_shape[1]
d_width = d_shape[1]
# Skip if there were no descriptors generated for this
# frame
if d_shape[1] == 0:
continue
video_info_mat_fps.append(ifp)
video_desc_mat_fps.append(dfp)
video_num_desc += d_shape[0]
# If combined descriptor height exceeds the per-item limit,
# generate a random subsample index list
ssi = None
if video_num_desc > per_item_limit:
ssi = sorted(
numpy.random.permutation(video_num_desc)[:per_item_limit]
)
video_num_desc = len(ssi)
r_map[uid] = (video_info_mat_fps, video_desc_mat_fps,
running_height, ssi)
running_height += video_num_desc