def test_overlapping_slice(cl):
input = NamedVideoStream(cl, 'test1')
frame = cl.io.Input([input])
slice_frame = cl.streams.Slice(frame,
partitions=[
cl.partitioner.strided_ranges(
[(0, 15), (5, 25), (15, 35)], 1)
])
sample_frame = cl.streams.Range(slice_frame,
ranges=[
SliceList([
{
'start': 0,
'end': 10
},
{
'start': 5,
'end': 15
},
{
'start': 5,
'end': 15
},
])
])
unsliced_frame = cl.streams.Unslice(sample_frame)
output = NamedStream(cl, 'test_slicing')
output_op = cl.io.Output(unsliced_frame, [output])
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
assert output.len() == 30
def test_bind_op_args(cl):
input = NamedVideoStream(cl, 'test1')
frame = cl.io.Input([input, input])
range_frame = cl.streams.Range(frame,
ranges=[{
'start': 0,
'end': 1
} for _ in range(2)])
test_out = cl.ops.TestPy(frame=range_frame,
kernel_arg=1,
x=[1, 10],
y=[5, 50])
outputs = [NamedStream(cl, 'test_hist_0'), NamedStream(cl, 'test_hist_1')]
output_op = cl.io.Output(test_out, outputs)
pairs = [(1, 5), (10, 50)]
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
for i, (x, y) in enumerate(pairs):
values = list(outputs[i].load())
p = values[0]
assert p['x'] == x
assert p['y'] == y
def test_python_stencil_batch_kernel(cl):
input = NamedVideoStream(cl, 'test1')
frame = cl.io.Input([input])
range_frame = cl.streams.Range(frame, ranges=[{'start': 0, 'end': 30}])
test_out = cl.ops.TestPyStencilBatch(frame=range_frame, batch=50)
output = NamedStream(cl, 'test_hist')
output_op = cl.io.Output(test_out, [output])
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
next(output.load())
def test_slice(cl):
input = NamedVideoStream(cl, 'test1')
frame = cl.io.Input([input])
slice_frame = cl.streams.Slice(frame, partitions=[cl.partitioner.all(50)])
unsliced_frame = cl.streams.Unslice(slice_frame)
output = NamedStream(cl, 'test_slicing')
output_op = cl.io.Output(unsliced_frame, [output])
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
assert input.len() == output.len()
def run(self, cl, device):
input = NamedVideoStream(cl, 'test1_inplace')
frame = cl.io.Input([input])
hist = cl.ops.Histogram(frame=frame, device=device)
output = NamedStream(cl, 'test_hist')
output_op = cl.io.Output(hist, [output])
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
next(output.load())
def test_unbounded_state(cl):
input = NamedVideoStream(cl, 'test1')
frame = cl.io.Input([input])
slice_frame = cl.streams.Slice(frame, partitions=[cl.partitioner.all(50)])
increment = cl.ops.TestIncrementUnbounded(ignore=slice_frame)
unsliced_increment = cl.streams.Unslice(increment)
output = NamedStream(cl, 'test_unbounded_state')
output_op = cl.io.Output(unsliced_increment, [output])
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
assert output.len() == input.len()
def test_wider_than_packet_stencil(cl):
input = NamedVideoStream(cl, 'test1')
frame = cl.io.Input([input])
sample_frame = cl.streams.Range(frame, ranges=[{'start': 0, 'end': 3}])
flow = cl.ops.OpticalFlow(frame=sample_frame, stencil=[0, 1])
output = NamedStream(cl, 'test_stencil')
output_op = cl.io.Output(flow, [output])
cl.run(output_op,
PerfParams.manual(1, 1, pipeline_instances_per_node=1),
cache_mode=CacheMode.Overwrite,
show_progress=False)
assert output.len() == 3
def get_maskrcnn_by_fid(sc, video_name, fid):
maskrcnn_stream = NamedStream(sc, video_name + '_maskrcnn')
seq = sc.sequence(maskrcnn_stream._name)
obj = seq.load(workers=1, rows=[fid])
metadata = next(obj)
if len(metadata) < 2:
print(fid, len(metadata))
return None, None
PERSON_CATEGORY = maskrcnn_detection.CATEGORIES.index('person')
# filter two player
bbox_size_list = []
for m in metadata:
if int(m['label']) == PERSON_CATEGORY and m['score'] > 0.9:
bbox_size_list += [m['bbox']['x2'] - m['bbox']['x1']]
else:
bbox_size_list += [0]
top2 = np.argsort(bbox_size_list)[-2:]
playerA, playerB = metadata[top2[0]], metadata[top2[1]]
if playerA['bbox']['y1'] > playerB['bbox']['y1']:
mask_fg = playerA['mask']
mask_bg = playerB['mask']
else:
mask_fg = playerB['mask']
mask_bg = playerA['mask']
return mask_fg, mask_bg
def test_slice_args(cl):
frame = cl.io.Input([NamedVideoStream(cl, 'test1')])
slice_frame = cl.streams.Slice(
frame, [cl.partitioner.ranges([[0, 1], [1, 2], [2, 3]])])
test = cl.ops.TestSliceArgs(frame=slice_frame,
arg=[SliceList([i for i in range(3)])])
unsliced_frame = cl.streams.Unslice(test)
output = NamedStream(cl, 'test_slicing')
output_op = cl.io.Output(unsliced_frame, [output])
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
num_rows = 0
list(output.load())
def run_spacer_job(spacer, spacing):
frame = cl.io.Input([NamedVideoStream(cl, 'test1')])
hist = cl.ops.Histogram(frame=frame)
space_hist = spacer(input=hist, spacings=[spacing])
output = NamedStream(cl, 'test_space')
output_op = cl.io.Output(space_hist, [output])
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
return output
def test_bounded_state(cl):
warmup = 3
frame = cl.io.Input([NamedVideoStream(cl, 'test1')])
increment = cl.ops.TestIncrementBounded(ignore=frame, bounded_state=warmup)
sampled_increment = cl.streams.Gather(increment,
indices=[[0, 10, 25, 26, 27]])
output = NamedStream(cl, 'test_bounded_state')
output_op = cl.io.Output(sampled_increment, [output])
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
num_rows = 0
expected_output = [0, warmup, warmup, warmup + 1, warmup + 2]
for buf in output.load():
(val, ) = struct.unpack('=q', buf)
assert val == expected_output[num_rows]
num_rows += 1
assert num_rows == 5
def get_densepose_by_fid(sc, video_name, fid):
def fid2idx(fid):
return frame_ids_dict[video_name].index(fid)
densepose_stream = NamedStream(sc, video_name[:-4] + '_densepose')
seq = sc.sequence(densepose_stream._name)
obj = seq.load(workers=1, rows=[fid2idx(fid)])
metadata = next(obj)
assert len(metadata) >= 2, "Player not detected!"
# filter two player by shoulder dist
# shoulder_dist = []
# for person in metadata:
# if person['keyp'][2, Person.LShoulder] < Person.KP_THRESH or person['keyp'][2, Person.RShoulder] < Person.KP_THRESH:
# shoulder_dist += [-1]
# else:
# shoulder_dist += [np.linalg.norm(person['keyp'][:2, Person.LShoulder] - person['keyp'][:2, Person.RShoulder])]
# top2 = np.argsort(shoulder_dist)[-2:]
# personA = Person(metadata[top2[0]]['bbox'], metadata[top2[0]]['keyp'], metadata[top2[0]]['mask'], score=metadata[top2[0]]['score'])
# personB = Person(metadata[top2[1]]['bbox'], metadata[top2[1]]['keyp'], metadata[top2[1]]['mask'], score=metadata[top2[1]]['score'])
# if personA.keyp[1, 5] >= personB.keyp[1, 5]:
# person_fg = personA
# person_bg = personB
# else:
# person_fg = personB
# person_bg = personA
# filter two player by bbox area
bbox_area = []
for person in metadata:
area = (person['bbox'][2] - person['bbox'][0]) * (person['bbox'][3] -
person['bbox'][1])
bbox_area.append(area)
top2 = np.argsort(bbox_area)[-2:]
person_fg = person_bg = None
personA = Person(metadata[top2[0]]['bbox'],
metadata[top2[0]]['keyp'],
metadata[top2[0]]['mask'],
score=metadata[top2[0]]['score'])
personB = Person(metadata[top2[1]]['bbox'],
metadata[top2[1]]['keyp'],
metadata[top2[1]]['mask'],
score=metadata[top2[1]]['score'])
if personA.keyp[1, Person.LShoulder] >= personB.keyp[1, Person.LShoulder]:
person_fg = personA
person_bg = personB
else:
person_fg = personB
person_bg = personA
return person_fg, person_bg
def test_auto_ingest(cl):
(vid1_path, vid2_path) = download_videos()
input = NamedVideoStream(cl, 'test3', path=vid1_path)
frame = cl.io.Input([input])
hist = cl.ops.Histogram(frame=frame)
output = NamedStream(cl, 'test_hist')
output_op = cl.io.Output(hist, [output])
cl.run(output_op,
PerfParams.estimate(),
cache_mode=CacheMode.Overwrite,
show_progress=False)
run(['rm', '-rf', vid1_path, vid2_path])
def test_fetch_resources(cl):
with tempfile.NamedTemporaryFile() as f:
f.write(b'0')
f.flush()
input = NamedVideoStream(cl, 'test1')
frame = cl.io.Input([input])
range_frame = cl.streams.Range(frame, ranges=[{'start': 0, 'end': 3}])
test_out = cl.ops.ResourceTest(frame=frame, path=f.name)
output = NamedStream(cl, 'test_hist')
output_op = cl.io.Output(test_out, [output])
cl.run(output_op,
PerfParams.estimate(pipeline_instances_per_node=2),
cache_mode=CacheMode.Overwrite,
show_progress=False)
def test_perf_params(cl):
frame = cl.io.Input([NamedVideoStream(cl, 'test1')])
hist = cl.ops.Histogram(frame=frame)
ghist = cl.streams.Gather(hist, [[0]])
output_op = cl.io.Output(ghist, [NamedStream(cl, '_ignore')])
cl.run(output_op,
PerfParams.manual(10, 10),
show_progress=False,
cache_mode=CacheMode.Overwrite)
cl.run(output_op,
PerfParams.estimate(),
show_progress=False,
cache_mode=CacheMode.Overwrite)
def test_no_workers(no_workers_cl):
cl = no_workers_cl
input = NamedVideoStream(cl, 'test1')
frame = cl.io.Input([input])
hist = cl.ops.Histogram(frame=frame)
output_op = cl.io.Output(hist, [NamedStream(cl, '_ignore')])
exc = False
try:
cl.run(output_op,
PerfParams.estimate(),
show_progress=False,
cache_mode=CacheMode.Overwrite)
except ScannerException:
exc = True
assert exc
def test_profiler(cl):
frame = cl.io.Input([NamedVideoStream(cl, 'test1')])
hist = cl.ops.Histogram(frame=frame)
ghist = cl.streams.Gather(hist, [[0]])
output_op = cl.io.Output(ghist, [NamedStream(cl, '_ignore')])
time_start = time.time()
job_id = cl.run(output_op,
PerfParams.estimate(),
show_progress=False,
cache_mode=CacheMode.Overwrite)
print('Time', time.time() - time_start)
profile = cl.get_profile(job_id)
f = tempfile.NamedTemporaryFile(delete=False, suffix='.trace')
f.close()
profile.write_trace(f.name)
profile.statistics()
run(['rm', '-f', f.name])
def test_pose(sc):
vid = [NamedVideoStream(sc, 'test1')]
frame = sc.io.Input(vid)
frame_sample = sc.streams.Gather(frame, [list(range(0, 1000, 100))])
pose = sc.ops.OpenPose(
frame=frame_sample,
device=DeviceType.GPU,
pose_num_scales=6,
pose_scale_gap=0.16,
compute_hands=True,
hand_num_scales=6,
hand_scale_gap=0.16,
compute_face=True,
batch=5
)
output = NamedStream(sc, 'test1-pose')
output_op = sc.io.Output(pose, [output])
sc.run(output_op, PerfParams.estimate())
def main():
print("Prepare videos and frames")
# video = Video.objects.filter(path__contains='men_single_final_gold')[0]
# movie_path = video.path
# movie_name = video.item_name()
sc = Client()
stride = 1
input_stream = NamedVideoStream(sc, movie_name, path=movie_path)
frame = sc.io.Input([input_stream])
strided_frame = sc.streams.Stride(frame, [stride])
print('Running Scanner MaskRCNN op')
maskrcnn_frame = sc.ops.MaskRCNNDetectObjects(
frame=strided_frame,
device=DeviceType.GPU, #if sc.has_gpu() else DeviceType.CPU,
batch=8,
confidence_threshold=0.5,
min_image_size=800)
maskrcnn_stream = NamedStream(sc, movie_name + '_maskrcnn')
output_op = sc.io.Output(maskrcnn_frame, [maskrcnn_stream])
sc.run(output_op,
scannerpy.common.PerfParams.estimate(),
cache_mode=scannerpy.CacheMode.Ignore)
print('Writing MaskRCNN metadata into frames')
drawn_frame = sc.ops.DrawMaskRCNN(frame=strided_frame,
bundled_data=sc.io.Input(
[maskrcnn_stream]),
min_score_thresh=0.5)
drawn_stream = NamedVideoStream(sc, movie_name + '_maskrcnn_draw')
output_op = sc.io.Output(drawn_frame, [drawn_stream])
sc.run(output_op,
scannerpy.common.PerfParams.estimate(),
cache_mode=scannerpy.CacheMode.Overwrite)
drawn_stream.save_mp4('/app/result/' + movie_name + '_maskrcnn')
def load(self, ty=None, fn=None, rows=None, workers=16):
"""
Loads the results of a Scanner computation into Python.
Kwargs:
fn: Optional function to apply to the binary blobs as they are read
in.
Returns:
Generator that yields either a numpy array for frame columns or
a binary blob for non-frame columns (optionally processed by the
`fn`).
"""
self._load_meta()
# If the column is a video, then dump the requested frames to disk as
# PNGs and return the decoded PNGs
if (self._descriptor.type == protobufs.Video
and self._video_descriptor.codec_type ==
protobufs.VideoDescriptor.H264):
png_table_name = self._sc._png_dump_prefix.format(
self._table.name(), self._name)
frame = self._sc.io.Input([NamedVideoStream(self._sc, self._table.name())])
enc_input = frame
if rows is not None:
sampled_frame = self._sc.streams.Gather(frame, indices=[rows])
enc_input = sampled_frame
img = self._sc.ops.ImageEncoder(frame=enc_input)
output = [NamedStream(self._sc, png_table_name)]
output_op = self._sc.io.Output(img, output)
self._sc.run(output_op, PerfParams.estimate(), cache_mode=CacheMode.Overwrite, show_progress=False)
return output[0].load()
elif self._descriptor.type == protobufs.Video:
frame_type = self._video_descriptor.frame_type
if frame_type == protobufs.U8:
dtype = np.uint8
elif frame_type == protobufs.F32:
dtype = np.float32
elif frame_type == protobufs.F64:
dtype = np.float64
def raw_frame_gen(shape0, shape1, shape2, typ):
def parser(bufs):
output = np.frombuffer(bufs, dtype=typ)
return output.reshape((shape0, shape1, shape2))
return parser
parser_fn = raw_frame_gen(
self._video_descriptor.height, self._video_descriptor.width,
self._video_descriptor.channels, dtype)
return self._load(fn=parser_fn, rows=rows, workers=workers)
else:
# Use a deserialize function if provided.
# If not, use a type if provided.
# If not, attempt to determine the type from the column's table descriptor.
# If that doesn't work, then assume no deserialization function, and return bytes.
if fn is None:
if ty is None:
type_name = self._descriptor.type_name
if type_name != "":
ty = scannertypes.get_type_info_cpp(type_name)
if ty is not None:
fn = ty.deserialize
return self._load(fn, rows=rows, workers=workers)
def main():
print("Prepare videos and frames")
video = Video.objects.filter(path__contains='wim')[0]
video_path = video.path
video_name = video.item_name()
sc = Client()
stride = 1
input_stream = NamedVideoStream(sc, video_name, path=video_path)
frame = sc.io.Input([input_stream])
# run on all frames
# running_frame = sc.streams.Stride(frame, [stride])
hit_annotation = pickle.load(
open('/app/data/pkl/hit_annotation_tennis.pkl',
'rb'))[video_name + '.mp4']
# run on selected frames
# hit_dict = []
# for h in hit_annotation.values():
# hit_dict += h
hit_dict = hit_annotation
# frame_ids = [i for point in hit_dict for i in range(point[0]['fid']-25, point[-1]['fid']+25) ]
frame_ids = [
i for point in hit_dict for i in range(point[0] - 25, point[-1] + 25)
]
frame_ids.sort()
running_frame = sc.streams.Gather(frame, [frame_ids])
print('Running Scanner DensePose op on %d frames' % (len(frame_ids)))
densepose_frame = sc.ops.DensePoseDetectPerson(frame=running_frame,
device=DeviceType.GPU,
batch=1,
confidence_threshold=0.5,
nms_threshold=0.2)
densepose_stream = NamedStream(sc, video_name + '_densepose')
output_op = sc.io.Output(densepose_frame, [densepose_stream])
sc.run(output_op,
scannerpy.common.PerfParams.estimate(),
cache_mode=scannerpy.CacheMode.Overwrite)
exit()
print('Writing DensePose metadata into frames')
drawn_frame = sc.ops.DrawDensePose(frame=running_frame,
bundled_data=sc.io.Input(
[densepose_stream]),
min_score_thresh=0.5,
show_body=True)
drawn_stream = NamedVideoStream(sc, video_name + '_densepose_draw_uvbody')
output_op = sc.io.Output(drawn_frame, [drawn_stream])
sc.run(output_op,
scannerpy.common.PerfParams.estimate(),
cache_mode=scannerpy.CacheMode.Overwrite)
drawn_stream.save_mp4('/app/result/' + video_name + '_densepose_uvbody')
drawn_frame = sc.ops.DrawDensePose(frame=running_frame,
bundled_data=sc.io.Input(
[densepose_stream]),
min_score_thresh=0.5,
show_body=False)
drawn_stream = NamedVideoStream(sc, video_name + '_densepose_draw_full')
output_op = sc.io.Output(drawn_frame, [drawn_stream])
sc.run(output_op,
scannerpy.common.PerfParams.estimate(),
cache_mode=scannerpy.CacheMode.Overwrite)
drawn_stream.save_mp4('/app/result/' + video_name + '_densepose_full')
def test_fault_tolerance(fault_cl):
force_kill_spawn_port = 5012
normal_spawn_port = 5013
def worker_killer_task(config, master_address):
from scannerpy import Config, start_worker, protobufs
import time
import grpc
import subprocess
import signal
import os
import scanner.metadata_pb2 as metadata_types
import scanner.engine.rpc_pb2 as rpc_types
import scanner.types_pb2 as misc_types
# Spawn a worker that we will force kill
script_dir = os.path.dirname(os.path.realpath(__file__))
with open(os.devnull, 'w') as fp:
p = subprocess.Popen([
'python3 ' + script_dir +
'/spawn_worker.py {:d}'.format(force_kill_spawn_port)
],
shell=True,
stdout=fp,
stderr=fp,
preexec_fn=os.setsid)
# Wait a bit for the worker to do its thing
time.sleep(10)
# Force kill worker process to trigger fault tolerance
os.killpg(os.getpgid(p.pid), signal.SIGTERM)
p.kill()
p.communicate()
# Wait for fault tolerance to kick in
time.sleep(15)
# Spawn the worker again
subprocess.call([
'python3 ' + script_dir +
'/spawn_worker.py {:d}'.format(normal_spawn_port)
],
shell=True)
master_addr = fault_cl._master_address
killer_process = Process(target=worker_killer_task,
args=(fault_cl.config, master_addr))
killer_process.daemon = True
killer_process.start()
input = NamedVideoStream(fault_cl, 'test1')
frame = fault_cl.io.Input([input])
range_frame = fault_cl.streams.Range(frame,
ranges=[{
'start': 0,
'end': 20
}])
sleep_frame = fault_cl.ops.SleepFrame(ignore=range_frame)
output = NamedStream(fault_cl, 'test_fault')
output_op = fault_cl.io.Output(sleep_frame, [output])
fault_cl.run(output_op,
PerfParams.estimate(pipeline_instances_per_node=1),
cache_mode=CacheMode.Overwrite,
show_progress=False)
assert output.len() == 20
# Shutdown the spawned worker
channel = grpc.insecure_channel('localhost:' + str(normal_spawn_port),
options=[('grpc.max_message_length',
24499183 * 2)])
worker = protobufs.WorkerStub(channel)
try:
worker.Shutdown(protobufs.Empty())
except grpc.RpcError as e:
status = e.code()
if status == grpc.StatusCode.UNAVAILABLE:
print('could not shutdown worker!')
exit(1)
else:
raise ScannerException(
'Worker errored with status: {}'.format(status))
killer_process.join()
