def main():
if not SERVER_URL:
print('please specify server host:port')
return
video_fname = "/data/dataset/jackson-town-square.mp4"
csv_in = "/data/dataset/jackson-town-square.csv"
avg_fname = "/data/dataset/jackson-town-square.npy"
num_frames = 500
start_frame = 1
data, nb_classes = dataPrep.get_data_for_test(
csv_in, video_fname, avg_fname,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=['car'],
resol=(50, 50))
X, Y = data
print(Y)
#r = requests.get(SERVER_URL + '/' + 'metadata')
TP = 0
TN = 0
FP = 0
FN = 0
for i in range(len(X)):
print("Sending request for %d frame" %(i))
payload = {
'signature_name': 'serving_default',
'inputs':
{
'image': [ X[i].tolist() ]
},
}
# print(payload)
r = requests.post(SERVER_URL + ':predict', json=payload)
probs = r.json()['outputs']
print('Confidence: %f'%(probs[0][1]))
if probs[0][1] > 0.5:
if Y[i][1] == 1.:
TP +=1
else:
FP +=1
else:
if Y[i][1] == 0.:
TN +=1
else:
FN +=1
print('Accuracy: %f' % ((TP+TN)/float(TP+FP+TN+FN)))
def main():
SERVER_URL = 'http://localhost:8502/v1/models/spm'
if not SERVER_URL:
print('please specify server host:port')
return
video_fname = "/data/dataset/jackson-town-square.mp4"
csv_in = "/data/dataset/jackson-town-square.csv"
avg_fname = "/data/dataset/jackson-town-square.npy"
num_frames = 500
start_frame = 1
data, nb_classes = dataPrep.get_data_for_test(
csv_in, video_fname, avg_fname,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=['car'],
resol=(50, 50))
X, Y = data
'''
for i in range(len(X)):
print("Sending request for %d frame" %(i))
payload = {
'signature_name': 'serving_default',
'inputs':
{
'image': [ X[i].tolist() ]
},
}
# print(payload)
r = requests.post(SERVER_URL + ':predict', json=payload)
probs = r.json()['outputs']
print('Confidence: %f'%(probs[0][1]))
'''
async def sendRequests(X,Y):
pass
loop = asyncio.get_event_loop()
loop.run_until_complete(sendRequests(X,Y))
def main():
video_fname = "/data/dataset/jackson-town-square.mp4"
csv_in = "/data/dataset/jackson-town-square.csv"
avg_fname = "/data/dataset/jackson-town-square.npy"
avg_fname_orig = "/data/dataset/jackson-town-square_400_600.npy"
num_frames = 30
start_frame = 3
target = 'car'
# Param: difference detector
dist_metric = 'mse'
thresh = 0.00024391713548410724 #0.002 #0.00024391713548410724
frame_delay = 1
feature_fn = RawImage.compute_feature
get_distance_fn = RawImage.get_distance_fn
# Param: specialized NN
c_low = 0.10
c_high = 0.90
# Param: time
ddTotalTime = 0
sNNTotalTime = 0
rNNTotalTime = 0
start = time.time()
data, nb_classes = dataPrep.get_data_for_test(csv_in,
video_fname,
avg_fname,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=[target],
resol=(50, 50),
shiftEn=1)
X, Y = data
print(X[0].shape)
origData, orig_nb_classes = dataPrep.get_data_for_test(
csv_in,
video_fname,
avg_fname_orig,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=[target],
resol=(400, 600),
shiftEn=0)
origX, origY = origData
print(origX[0].shape)
# start = time.time()
ddTotalTime = ddTotalTime + (time.time() - start)
conf = None
ddCount = 0
sNNFrameCount = 0
rNNFrameCount = 0
for i in range(1):
if i == 0:
sNNFrameCount, rNNFrameCount, sNNTotalTime, rNNTotalTime, conf = runNNRequest(
sNNFrameCount, rNNFrameCount, sNNTotalTime, rNNTotalTime, X[i],
origX[i], c_low, c_high, target)
print('Frame %d: Confidence = %f' % (i, conf))
print("0. Total frames: {} from frame {} to frame {}".format(
len(X), start_frame, start_frame + num_frames - 1))
print("1. Difference detector: {} times, {:.2f} sec".format(
ddCount, ddTotalTime))
print("2. Specialized NN: {} times, {:.2f} sec".format(
sNNFrameCount, sNNTotalTime))
print("3. Reference NN: {} times, {:.2f} sec".format(
rNNFrameCount, rNNTotalTime))
def main():
if not SERVER_URL:
print('please specify server host:port')
return
video_fname = "/data/dataset/jackson-town-square.mp4"
csv_in = "/data/dataset/jackson-town-square.csv"
avg_fname = "/data/dataset/jackson-town-square.npy"
num_frames = 500
start_frame = 1
# Param: difference detector
dist_metric = 'mse'
thresh = 0.002 #0.00024391713548410724
frame_delay = 1
feature_fn = RawImage.compute_feature
get_distance_fn = RawImage.get_distance_fn
data, nb_classes = dataPrep.get_data_for_test(csv_in,
video_fname,
avg_fname,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=['car'],
resol=(50, 50))
X, Y = data
#r = requests.get(SERVER_URL + '/' + 'metadata')
TP = 0
TN = 0
FP = 0
FN = 0
r = None
posCount = 0
for i in range(len(X)):
print("Sending request for %d frame" % (i))
payload = {
'signature_name': 'serving_default',
'inputs': {
'image': [X[i].tolist()]
},
}
# print(payload)
if i == 0:
r = requests.post(SERVER_URL + ':predict', json=payload)
else: # Difference detector start working from the 2nd frame
features = get_features(feature_fn, X[i - 1:i + 1])
dists = get_distances(get_distance_fn(dist_metric), features,
frame_delay)
if dists > thresh:
posCount = posCount + 1
r = requests.post(SERVER_URL + ':predict', json=payload)
probs = r.json()['outputs']
print('Confidence: %f' % (probs[0][1]))
if probs[0][1] > 0.5:
if Y[i][1] == 1.:
TP += 1
else:
FP += 1
else:
if Y[i][1] == 0.:
TN += 1
else:
FN += 1
print("posCount = {}".format(posCount))
print('Accuracy: %f' % ((TP + TN) / float(TP + FP + TN + FN)))
def main(num_of_frames=50, num_of_workers=4):
video_fname = "/data/dataset/jackson-town-square.mp4"
csv_in = "/data/dataset/jackson-town-square.csv"
avg_fname = "/data/dataset/jackson-town-square.npy"
avg_fname_orig = "/data/dataset/jackson-town-square_400_600.npy"
num_frames = num_of_frames
start_frame = 5001
target = 'car'
# Param: difference detector
dist_metric = 'mse'
thresh = 0.00024391713548410724
frame_delay = 1
feature_fn = RawImage.compute_feature
get_distance_fn = RawImage.get_distance_fn
# Param: specialized NN
c_low = 0.1
c_high = 0.9
# start = time.time()
data, nb_classes = dataPrep.get_data_for_test(csv_in,
video_fname,
avg_fname,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=[target],
resol=(50, 50),
shiftEn=1)
X, Y = data
origData, orig_nb_classes = dataPrep.get_data_for_test(
csv_in,
video_fname,
avg_fname_orig,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=[target],
resol=(608, 608),
shiftEn=0)
origX, origY = origData
meta_f = open('yolo.meta', 'r')
metadata = json.loads(meta_f.read())
# Run Difference Detector, and generate index of frames which need Specialized NN
start = time.time()
todos = [0]
for i in range(1, len(X)):
features = get_features(feature_fn, X[i - 1:i + 1])
dists = get_distances(get_distance_fn(dist_metric), features,
frame_delay)
if dists > thresh:
todos.append(i)
X_new = X[todos]
origX_new = origX[todos]
'''
tasks = []
for i in range(len(X_new)):
task = asyncio.ensure_future(sendNNRequest(metadata, X_new[i:i+1], origX_new[i:i+1], c_low, c_high, target))
tasks.append(task)
loop = asyncio.get_event_loop()
responses = loop.run_until_complete(asyncio.gather(*tasks))
'''
'''
executor = concurrent.futures.ThreadPoolExecutor(
max_workers=10,
)
'''
executor = concurrent.futures.ProcessPoolExecutor(
max_workers=num_of_workers, )
loop = asyncio.get_event_loop()
loop.run_until_complete(
sendAllRequests(executor, metadata, X_new, origX_new, c_low, c_high,
target))
print('Num of RNN: ', str(num_of_RNN), " Run Time: ", time.time() - start)
def main():
video_fname = "/data/dataset/jackson-town-square.mp4"
csv_in = "/data/dataset/jackson-town-square.csv"
avg_fname = "/data/dataset/jackson-town-square.npy"
avg_fname_orig = "/data/dataset/jackson-town-square_400_600.npy"
num_frames = 500
start_frame = 5001
target = 'car'
# Param: difference detector
dist_metric = 'mse'
thresh = 0.00024391713548410724 #0.002 #0.00024391713548410724
frame_delay = 1
feature_fn = RawImage.compute_feature
get_distance_fn = RawImage.get_distance_fn
# Param: specialized NN
c_low = 0.10
c_high = 0.90
# Param: time
ddTotalTime = 0
sNNTotalTime = 0
rNNTotalTime = 0
start = time.time()
data, nb_classes = dataPrep.get_data_for_test(
csv_in, video_fname, avg_fname,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=[target],
resol=(50, 50), shiftEn=1)
X, Y = data
print(X[0].shape)
origData, orig_nb_classes = dataPrep.get_data_for_test(
csv_in, video_fname, avg_fname_orig,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=[target],
resol=(400, 600), shiftEn=0)
origX, origY = origData
print(origX[0].shape)
ddTotalTime = ddTotalTime + (time.time() - start)
TP = 0
TN = 0
FP = 0
FN = 0
conf = None
ddCount = 0
sNNFrameCount = 0
rNNFrameCount = 0
for i in range(len(X)):
if i == 0:
sNNFrameCount, rNNFrameCount, sNNTotalTime, rNNTotalTime, conf = runNNRequest(
sNNFrameCount, rNNFrameCount, sNNTotalTime, rNNTotalTime, X[i], origX[i], c_low, c_high, target)
else: # Difference detector start working from the 2nd frame
start = time.time()
ddCount = ddCount + 1
features = get_features(feature_fn, X[i-1:i+1])
dists = get_distances(get_distance_fn(dist_metric), features, frame_delay)
ddTotalTime = ddTotalTime + (time.time() - start)
if dists > thresh:
sNNFrameCount, rNNFrameCount, sNNTotalTime, rNNTotalTime, conf = runNNRequest(
sNNFrameCount, rNNFrameCount, sNNTotalTime, rNNTotalTime, X[i], origX[i], c_low, c_high, target)
print('Frame %d: Confidence = %f'%(i, conf))
if conf > 0.5:
if Y[i][1] == 1.:
TP +=1
else:
FP +=1
else:
if Y[i][1] == 0.:
TN +=1
else:
FN +=1
print("0. Total frames: {} from frame {} to frame {}".format(len(X), start_frame, start_frame+num_frames-1))
print("1. Difference detector: {} times, {:.2f} sec".format(ddCount, ddTotalTime))
print("2. Specialized NN: {} times, {:.2f} sec".format(sNNFrameCount, sNNTotalTime))
print("3. Reference NN: {} times, {:.2f} sec".format(rNNFrameCount, rNNTotalTime))
print('Accuracy: %f' % ((TP+TN)/float(TP+FP+TN+FN)))
def main():
video_fname = "/data/dataset/jackson-town-square.mp4"
csv_in = "/data/dataset/jackson-town-square.csv"
avg_fname = "/data/dataset/jackson-town-square.npy"
avg_fname_orig = "/data/dataset/jackson-town-square_400_600.npy"
num_frames = 2000
start_frame = 5001
target = 'car'
# Param: difference detector
dist_metric = 'mse'
thresh = 0.00024391713548410724
frame_delay = 1
feature_fn = RawImage.compute_feature
get_distance_fn = RawImage.get_distance_fn
# Param: specialized NN
c_low = 0.1
c_high = 0.9
data, nb_classes = dataPrep.get_data_for_test(csv_in,
video_fname,
avg_fname,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=[target],
resol=(50, 50),
shiftEn=1)
X, Y = data
origData, orig_nb_classes = dataPrep.get_data_for_test(
csv_in,
video_fname,
avg_fname_orig,
num_frames=num_frames,
start_frame=start_frame,
OBJECTS=[target],
resol=(608, 608),
shiftEn=0)
origX, origY = origData
meta_f = open('yolo.meta', 'r')
metadata = json.loads(meta_f.read())
# Run Difference Detector, and generate index of frames which need Specialized NN
start = time.time()
todos = [0]
for i in range(1, len(X)):
features = get_features(feature_fn, X[i - 1:i + 1])
dists = get_distances(get_distance_fn(dist_metric), features,
frame_delay)
if dists > thresh:
todos.append(i)
X_new = X[todos]
origX_new = origX[todos]
loop = asyncio.get_event_loop()
results = loop.run_until_complete(
sendAllRequests(metadata, X_new, origX_new, c_low, c_high, target))
print(time.time() - start)