def main(args):
log_handler(LOGGER)
LOGGER.info(args)
gpu_candidate = args.gpus.split(',')
# prepare arguments
root_path = Path(args.root)
outdir = Path(args.outdir)
video_paths = list(root_path.glob('./**/*.avi'))
gpu_queue = Queue()
for i in gpu_candidate:
gpu_queue.put(i)
# multithreading
th_jobs = []
all_th_args = zip([gpu_queue]*len(video_paths), \
video_paths, \
[outdir]*len(video_paths), \
[args.option]*len(video_paths))
for th_args in all_th_args:
job = Thread(target=demo_one_video, args=th_args)
th_jobs.append(job)
job.start()
for j in th_jobs:
j.join()
LOGGER.info('Complete multithread processing with pipeline ')
def main(args: argparse.Namespace):
"""an interface to activate pyqt5 app"""
logger = logging.getLogger(__name__)
log_handler(logger)
logger.info(args)
with open(args.config, 'r') as config_file:
config = yaml.load(config_file)
video_path = Path(args.video)
output_path = Path('outputs')
if not output_path.exists():
output_path.mkdir(parents=True)
label_path = output_path / '{}_label.csv'.format(video_path.stem)
label_path = output_path / str(Path(
args.output)) if args.output else label_path
if not label_path.parent.exists():
label_path.parent.mkdir(parents=True)
app = QApplication(sys.argv)
video_app = VideoApp(args.video, str(label_path), **config)
try:
log_handler(video_app.logger)
app.exec()
except Exception as e:
logger.exception(e)
def main(args):
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpus
logger = logging.getLogger(__name__)
log_handler(logger, logging.getLogger('src.utils'))
logger.info(args)
# prepare data
X_test = list(Path(args.test).glob('*/*'))
y_true = [path.parent.name for path in X_test]
y_true = list(map(int, y_true))
y_true = np_utils.to_categorical(y_true)
X_test = np.array([
np.array(Image.open(i).resize(args.input_shape), dtype='float32')
for i in X_test
])
X_test = X_test / 255
logger.info('X_test shape={}, Y_test shape={}'.format(
X_test.shape, y_true.shape))
# models
for model_path in args.models:
K.clear_session()
if args.focal_loss:
model = load_model(
model_path,
custom_objects={'focal_loss_fixed': focal_loss(2, 2)})
else:
model = load_model(model_path)
# predict
y_pred = model.predict(X_test)
savepath_y_pred = str(Path(model_path).with_name('y_pred.npy'))
np.save(savepath_y_pred, y_pred)
# metrics
accuracy = top_n_accuracy(y_pred, y_true, 1)
cnf_matrix = confusion_matrix(y_true.argmax(axis=1),
y_pred.argmax(axis=1))
savepath_cnf_matrix = str(
Path(model_path).with_name('confusion_matrix.jpg'))
plot_confusion_matrix(
cnf_matrix,
classes=list('OX=A'),
to_img=savepath_cnf_matrix,
normalize=True,
title='Confusion matrix (acc={:.4f})'.format(accuracy))
logger.info('model: {}'.format(model_path))
logger.info(' - top-1-accuracy: {:.4f}'.format(accuracy))
logger.info(
' - save confusion matrix at {}'.format(savepath_cnf_matrix))
logger.info(' - save y_pred at {}'.format(savepath_y_pred))
print(
classification_report(y_true.argmax(axis=1),
y_pred.argmax(axis=1),
target_names=list('OX=A')))
def main(args):
logger = logging.getLogger(__name__)
log_handler(logger, logging.getLogger('src.visualize'))
logger.info(args)
record_path = Path(args.record)
save_video_path = record_path.parent / f'{record_path.stem}_{args.options}.avi'
save_video_path = save_video_path if args.save_video else None
with open(args.config, 'r') as f:
config = json.load(f)['outputs']
mouse_contours = None
if args.mouse_contours:
with open(args.mouse_contours, 'r') as f:
mouse_contours = json.load(f)
if args.save_video:
if not save_video_path.parent.exists():
save_video_path.parent.mkdir(parents=True)
save_video_path = str(save_video_path)
if args.options in ['detection', 'classification']:
with open(args.record, 'r') as f:
records = f.readlines()
records = [eval(line) for line in records]
df = convert_to_dataframe(records, args.options)
show_and_save_video(args.video,
df,
config,
from_=args.from_idx,
show_video=args.show_video,
save_video=save_video_path,
pause_flag=args.pause)
elif args.options == 'track':
df = pd.read_csv(args.record)
show_and_save_video(args.video,
df,
config,
from_=args.from_idx,
mouse_contours=mouse_contours,
show_video=args.show_video,
save_video=save_video_path,
pause_flag=args.pause)
elif args.options == 'action':
df = pd.read_csv(args.record)
if Path(args.record).name != 'fixed_action_paths.csv':
df = convert_to_dataframe(df, args.options)
df.to_csv(Path(args.record).with_name('fixed_action_paths.csv'))
show_and_save_video(args.video,
df,
config,
from_=args.from_idx,
show_video=args.show_video,
save_video=save_video_path,
pause_flag=args.pause)
def main(args: argparse.Namespace):
# setting log
out_dir_path = Path(args.outdir)
log_name = f"{datetime.now().strftime('%m%dT%H%M%S')}-{datetime.now().microsecond}.log"
log_path = out_dir_path / log_name
if not out_dir_path.exists():
out_dir_path.mkdir(parents=True)
logger = logging.getLogger(__name__)
log_handler(logger, logname=log_path if args.islog else None)
logger.info(args)
# run pyqt5
app = QApplication(sys.argv)
tracker = KalmanFilterTracker()
log_handler(tracker.logger)
app.exec()
def main(args: argparse.Namespace):
"""optical flow interface
Arguments:
args {argparse.Namespace} -- parameter parse fomr terminal
"""
# config
logger = logging.getLogger(__name__)
log_handler(logger)
logger.info(args)
with open(CONFIG_FILE) as config_file:
config = yaml.load(config_file)
# demo code for optical flow testing
with Video(args.video) as video:
prev_frame = video.read_frame(0)
prev_gray = cv2.cvtColor(prev_frame, cv2.COLOR_BGR2GRAY)
# limit the feature area
pos_0 = cv2.goodFeaturesToTrack(prev_gray[:300, :400],
mask=None,
**config['shitomasi'])
mask = np.zeros_like(prev_frame)
while True:
frame_img = video.read_frame(video.frame_idx + 1)
frame_gray = cv2.cvtColor(frame_img, cv2.COLOR_BGR2GRAY)
# calc optical flow
pos_1, status, err = cv2.calcOpticalFlowPyrLK(prev_gray, frame_gray, \
pos_0, None, **config['lk'])
good_new = pos_1[status == 1]
good_old = pos_0[status == 1]
for new, old in zip(good_new, good_old):
pt1, pt2 = tuple(new.ravel()), tuple(old.ravel())
mask = cv2.line(mask, pt1, pt2, (0, 255, 0), 2)
frame_img = cv2.circle(frame_img, pt1, 5, (0, 0, 255), -1)
img = cv2.add(frame_img, mask)
cv2.imshow('frame', img)
k = cv2.waitKey(1)
if k in [27, ord('q')]:
cv2.destroyAllWindows()
break
prev_gray = frame_gray.copy()
pos_0 = good_new.reshape(-1, 1, 2)
def main(args):
log_handler(LOGGER, logging.getLogger('src.utils'))
LOGGER.info(args)
# prepare data
func_set = {
'avg': ensemble_avg,
'vote': ensemble_vote
}
X_test = list(Path(args.ground_truth).glob('*/*'))
y_true = [path.parent.name for path in X_test]
y_true = np.array(list(map(int, y_true)))
if args.predict_all:
y_preds = Path(args.predictions_dir).glob('*/y_pred.npy')
y_preds = list(map(str, y_preds))
for combine_length in range(2, len(y_preds)):
_checkpoint = (None, 0)
for combine_set in combinations(y_preds, combine_length):
local_y_preds = np.array([np.load(i) for i in combine_set])
local_y_pred = func_set.get(args.method)(local_y_preds)
accuracy = accuracy_score(local_y_pred, y_true, 1)
if _checkpoint[1] < accuracy:
_checkpoint = (combine_set, accuracy)
print('Limit len {:02d} - combination={}, acc={:.4f}'.format(
combine_length, sorted(map(lambda x: x.split('/')[1], _checkpoint[0])), _checkpoint[1]
))
else:
y_preds = np.array([np.load(i) for i in args.predictions])
LOGGER.info('predictions shape - {}'.format(y_preds.shape))
y_pred = func_set.get(args.method)(y_preds)
LOGGER.info('ensemble predictions shape - {}'.format(y_pred.shape))
# metrics
accuracy = accuracy_score(y_pred, y_true, 1)
cnf_matrix = confusion_matrix(y_true, y_pred)
savepath_cnf_matrix = str(Path('.') / 'ensemble_confusion_matrix.jpg')
plot_confusion_matrix(cnf_matrix,
classes=list('OX=A'),
to_img=savepath_cnf_matrix,
normalize=True,
title='Confusion matrix (acc={:.4f})'.format(accuracy))
LOGGER.info('Ensemble top-1-accuracy - {:.4f}'.format(accuracy))
print(classification_report(y_true, y_pred, target_names=list('OX=A')))
def main(args):
logger = logging.getLogger(__name__)
log_handler(logger)
logger.info(args)
outdir = Path(__file__).parent / '../output/path'
if not outdir.exists():
logger.info('{} not found, created.'.format(str(outdir)))
outdir.mkdir(parents=True)
observer_filepath = Path(args.observer)
paths_filepath = Path(args.paths)
fixed_observer_savepath = paths_filepath.parent / 'fixed_observer.csv'
# get fps from video
cap = cv2.VideoCapture(args.video)
fps = cap.get(cv2.CAP_PROP_FPS)
cap.release()
logger.info('video fps = {}'.format(fps))
# clean observer file
df_observer = pd.read_csv(observer_filepath)
df_observer = df_observer[[
'Time_Relative_hmsf', 'Duration_sf', 'Subject', 'Behavior',
'Modifier_1', 'Event_Type'
]]
df_observer = df_observer.dropna(subset=['Modifier_1'])
df_observer.columns = [ \
'timestamp_hmsf', 'duration', 'subject', \
'behavior', 'target', 'event_type']
df_observer['label'] = df_observer.apply(
lambda row: row['subject'].split(' ')[-1], axis=1)
df_observer['timestamp_hmsf'] = pd.to_datetime(
df_observer['timestamp_hmsf'])
df_observer['timestamp_hmsf'] = [
t.time() for t in df_observer['timestamp_hmsf']
]
df_observer.to_csv(fixed_observer_savepath, index=False)
logger.info('Complete to clean and save observer file at {}'.format(
fixed_observer_savepath))
# padding the records in duration
df_padding_event = df_observer.copy().reset_index(drop=True)
df_padding_event['frame_idx'] = df_padding_event.apply( \
lambda x: hmsf_to_idx(x['timestamp_hmsf'], fps), axis=1)
for index, row in df_observer.iterrows():
if row['event_type'] == 'State start':
logger.info('From row {:02} - {}'.format(index, row.tolist()))
for end_index, end_row in df_observer.loc[index + 1:, ].iterrows():
check_end_row = end_row['subject':'target'].tolist()
check_row = row['subject':'target'].tolist()
check_same_event = (check_end_row == check_row)
check_same_event = check_same_event and end_row[
'event_type'] == 'State stop'
if check_same_event:
# create the DaaFrame to pad the loss record per event
# date_range build wiht milliseconds
time1 = datetime.datetime.combine(datetime.date.min,
row['timestamp_hmsf'])
time2 = datetime.datetime.combine(
datetime.date.min, end_row['timestamp_hmsf'])
delta = time2 - time1
ts = pd.date_range(str(row['timestamp_hmsf']),
periods=delta.total_seconds() * 1000,
freq='L',
closed='right')
ts = [t.time() for t in ts.tolist()]
ts_data = {
'timestamp_hmsf': ts,
'duration': [0.0] * len(ts),
'subject': [row['subject']] * len(ts),
'behavior': [row['behavior']] * len(ts),
'target': [row['target']] * len(ts),
'event_type': ['State processing'] * len(ts),
'label': [row['label']] * len(ts)
}
df_ts = pd.DataFrame(data=ts_data)
df_ts['frame_idx'] = df_ts.apply( \
lambda x: hmsf_to_idx(x['timestamp_hmsf'], fps), axis=1)
df_ts.drop_duplicates(subset=['frame_idx'],
keep='first',
inplace=True)
df_padding_event = pd.concat([df_padding_event, df_ts],
sort=False)
logger.info('To row {:02} - {}\n{}'.format(
end_index, end_row.tolist(), '-' * 87))
break
df_padding_event.reset_index(drop=True, inplace=True)
df_padding_event.drop_duplicates(subset=['frame_idx', 'label'],
keep='first',
inplace=True)
logger.info('Complete to create the padding dataframe, shape={}'.format(
df_padding_event.shape))
# join action and path
# observer precise on millisecond, timestamp_ms should be remove the microsecond part
df_paths = pd.read_csv(paths_filepath)
df_final = pd.merge(df_paths,
df_padding_event,
how='left',
on=['frame_idx', 'label'])
df_final = add_group_id(df_final,
'block_idx',
'behavior',
'target',
gid_colname='action_idx')
logger.info('paths file shape={}'.format(df_paths.shape))
logger.info('Compete merge to final dataframe, shape={}'.format(
df_final.shape))
# save final result
action_path_filepath = paths_filepath.parent / 'action_paths.csv'
df_final.to_csv(action_path_filepath, index=False)
def main(args):
logdir = Path('logs')
outdir = Path(args.outdir)
trackpath_dir = outdir / Path(args.video).stem
if not logdir.exists():
logdir.mkdir(parents=True)
if not trackpath_dir.exists():
trackpath_dir.mkdir(parents=True)
logger = logging.getLogger(__name__)
log_handler(logger,
*LOGGERS,
logname=str(logdir / args.log) if args.log else None)
logger.info(args)
trackflow = build_flow(args.video, args.input, args.config)
# get timestamp and save path to file
video_filename = str(Path(args.video).stem)
cap = cv2.VideoCapture(args.video)
fps = cap.get(cv2.CAP_PROP_FPS)
total_frames = cap.get(cv2.CAP_PROP_FRAME_COUNT)
cap.release()
trackflow_all_label = []
label_col = sorted(trackflow.ref_keys)
for label, flow in trackflow.paths.items():
logger.info('get timestamp and convert {}'.format(label))
for bbox in tqdm(flow):
bbox.timestamp = bbox.frame_idx / fps * 1000
label_score = [
bbox.multiclass_result.get(l, None) for l in label_col
]
trackflow_all_label.append([
video_filename, fps, total_frames, bbox.frame_idx,
bbox.block_id, bbox.timestamp, label, bbox.confidence,
*label_score, bbox.block_confidence, *bbox.pt1, *bbox.pt2,
*bbox.center, *bbox_behavior_encoding(bbox.behavior)
])
trackflow_all_label = sorted(trackflow_all_label, key=lambda x: x[0])
label_col = ['{}_score'.format(l) for l in label_col]
trackflow_cols = [
'video_name', 'video_fps', 'video_nframes', 'frame_idx', 'block_idx',
'timestamp_ms', 'label', 'detect_score', *label_col, 'block_score',
'pt1.x', 'pt1.y', 'pt2.x', 'pt2.y', 'center.x', 'center.y', 'on_mouse'
]
path_savepath = str(trackpath_dir / 'paths.csv')
df_paths = pd.DataFrame(trackflow_all_label, columns=trackflow_cols)
df_paths.to_csv(path_savepath, index=False)
# save mouse contours
mouse_filename = '{}_mouse.json'.format(Path(args.video).stem)
mouse_cnts_filepath = str(trackpath_dir / mouse_filename)
mouse_cnts = {
str(m.frame_idx): {
'contour': m.contour.tolist(),
'contour_extend': m.contour_extend.tolist(),
'contour_extend_kernel': m.contour_extend_kernel.tolist(),
'center': m.center.tolist()
}
for m in trackflow.mouse_cnts
}
with open(mouse_cnts_filepath, 'w+') as f:
json.dump(mouse_cnts, f)
# show and/or save video
video_savepath = None
if args.save_video:
video_savepath = str(trackpath_dir / args.outvideo)
if args.show_video or args.save_video:
with open(args.config, 'r') as f:
config = json.load(f)
show_and_save_video(args.video,
df_paths,
config['outputs'],
from_=args.from_idx,
show_video=args.show_video,
save_video=video_savepath,
pause_flag=args.pause)
def main(args):
# preprocess
os.environ['CUDA_VISIBLE_DEVICES'] = ','.join(args.gpus)
outdir = Path('outputs') / args.name
if not outdir.exists():
outdir.mkdir(parents=True)
log_handler(*LOGGERS, logname=str(outdir / 'log.txt'))
if Path(args.config).exists():
with open(args.config, 'r') as f:
config = json.load(f)
LOGGER.info(args)
# build and compile model
K.clear_session()
model = get_model(args, config)
model.model = multi_gpu_model(model.model, gpus=len(
args.gpus)) if len(args.gpus) > 1 else model.model
optimizer = get_optimizer(config[args.backend]['optimizer'],
lr=config[args.backend]['lr'])
loss = focal_loss(gamma=2, alpha=2)
model.model.compile(loss=[loss],
optimizer=optimizer,
metrics=[config[args.backend]['metrics']])
estimate_gbytes = get_model_memory_usage(config[args.backend]['bz'],
model.model)
model.model.summary(print_fn=lambda x: LOGGER.info(x + '\n'))
LOGGER.info(
'Estimate model required {} gbytes GPU memory'.format(estimate_gbytes))
# preprocess generator
_data_gen_params = {
'target_size': config[args.backend]['keras']['input_shape'][:2],
'batch_size': config[args.backend]['bz'],
'class_mode': 'categorical',
'shuffle': True
}
count_train_data = len(list(Path(args.train).glob('**/*')))
count_test_data = len(list(Path(args.test).glob('**/*')))
train_data_aug = ImageDataGenerator(**config['imgaug']['train'])
test_data_aug = ImageDataGenerator(**config['imgaug']['test'])
train_batch = train_data_aug.flow_from_directory(args.train,
**_data_gen_params)
test_batch = test_data_aug.flow_from_directory(args.test,
**_data_gen_params)
train_data_gen = normalize_generator(train_batch)
test_data_gen = normalize_generator(test_batch)
LOGGER.info(
'Complete generator preprocess with {} traing data and {} test data'.
format(count_train_data, count_test_data))
# callbacks
model_savepath = str(outdir / '{}.h5'.format(args.backend))
checkpoint = ModelCheckpoint(model_savepath,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min')
earlystop = EarlyStopping(monitor='val_loss',
min_delta=0,
patience=25,
verbose=1,
mode='auto')
reducelr = ReduceLROnPlateau(monitor='val_loss',
factor=0.8,
patience=5,
min_lr=1e-10,
verbose=1)
tensorboard = TensorBoard(log_dir=str(outdir / 'tensorboard'),
write_grads=True,
write_images=True,
write_graph=False)
LOGGER.info('Complete callbacks declarement')
# train
history = model.model.fit_generator(
train_data_gen,
steps_per_epoch=count_train_data // config[args.backend]['bz'],
epochs=config[args.backend]['epochs'],
validation_data=test_data_gen,
validation_steps=count_test_data // config[args.backend]['bz'],
callbacks=[checkpoint, earlystop, reducelr, tensorboard])
# save model config
history_dataframe = pd.DataFrame(history.history)
history_dataframe.to_csv(str(outdir / 'history.csv'), index=False)
record_config = model.config
record_config['imgaug'] = config['imgaug']
if args.outimg:
plot_model(model.model, to_file=str(outdir / 'model_structure.jpg'))
with open(str(outdir.parent / 'README.txt'), 'a+') as f:
f.write('{}\t-\t{}\n'.format(args.name, args.comment))
with open(str(outdir / 'config.json'), 'w+') as f:
json.dump(record_config, f, indent=4)
def main(args: argparse.Namespace):
"""[summary]
an interface to handle input and parse to different method
Arguments:
args {argparse.Namespace} -- parameter parse fomr terminal
"""
# config
logger = logging.getLogger(__name__)
log_handler(logger, logging.getLogger(Video.__class__.__name__))
logger.info(args)
with open(CONFIG_FILE) as config_file:
config = yaml.load(config_file)
# test single frame from video and the conventional find contour method
with Video(args.input) as video:
# multiprocessing calc the contour
pending_frame_idx = list(range(0, video.frame_count, config['general']['skip_per_nframe']))
logger.info('process pending frame index: %d', len(pending_frame_idx))
logger.info('cpu count: %d (will used %d)', cpu_count(), (cpu_count()*3//4))
with Pool(processes=(cpu_count()*3//4)) as pool:
# basic contours
mp_args = zip([args.input]*len(pending_frame_idx), \
pending_frame_idx, \
[config]*len(pending_frame_idx), \
[args.preprocess]*len(pending_frame_idx), \
[args.option]*len(pending_frame_idx))
mp_targets = pool.starmap_async(get_contour_from_video, mp_args)
mp_targets = mp_targets.get()
mp_targets = [DetectionTarget(*i) for i in mp_targets if i]
mp_targets = sorted(mp_targets, key=lambda x: x.frame_idx)
_basic_target_counts = len(mp_targets)
# interpolate contours
_check_frame_len = 0
while True:
# find the new pending index
pair_targets = [(mp_targets[i], mp_targets[i+1]) for i in range(len(mp_targets)-1)]
check_frame_idx = [i.is_shifting(j, config['general']['tolerable_shifting_dist']) \
for i, j in pair_targets]
if len(check_frame_idx) == _check_frame_len:
break
_check_frame_len = len(check_frame_idx)
logger.info('check_frame_idx len=%d (all=%d), any=%s', len(check_frame_idx), video.frame_count, any(check_frame_idx))
if not any(check_frame_idx):
break
pending_frame_idx = list(compress(pair_targets, check_frame_idx))
pending_frame_idx = [(i.frame_idx+j.frame_idx)//2 \
for i, j in pending_frame_idx if i.frame_idx+1 < j.frame_idx]
# multiprocessing calc
mp_args = zip([args.input]*len(pending_frame_idx), \
pending_frame_idx, \
[config]*len(pending_frame_idx), \
[args.preprocess]*len(pending_frame_idx), \
[args.option]*len(pending_frame_idx))
mp_interpolate_targets = pool.starmap_async(get_contour_from_video, mp_args)
mp_interpolate_targets = mp_interpolate_targets.get()
mp_targets += [DetectionTarget(*i) for i in mp_interpolate_targets if i]
mp_targets = sorted(mp_targets, key=lambda x: x.frame_idx)
logger.info('#contours after interpolate: %d -> %d', \
_basic_target_counts, len(mp_targets))
video.detect_targets += mp_targets
video.save(args.savepath, draw_cnts=True)
def main(args):
logger = logging.getLogger(__name__)
log_handler(logger)
logger.info(args)
# setting GPU
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpus
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.5
K.tensorflow_backend.set_session(tf.Session(config=config))
# prepare video and detection results
video_path = Path(args.video)
cap = cv2.VideoCapture(str(video_path))
with open(args.input, 'r') as f:
bbox_results = f.readlines()
bbox_results = [eval(b.rstrip('\n')) for b in bbox_results]
# prepare model predict results
y_preds = {}
for model_path in args.models:
logger.info('model - {}'.format(model_path))
K.clear_session()
model = load_model(
model_path, custom_objects={'focal_loss_fixed': focal_loss(2, 2)})
y_preds[model_path] = []
for record in tqdm(bbox_results):
frame_idx, bboxes = record
model_predict_bbox = deepcopy(bboxes)
cap.set(1, frame_idx - 1)
success, frame = cap.read()
if success:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
for bbox_idx, bbox in enumerate(model_predict_bbox):
y1, x1, y2, x2 = [int(i) for i in bbox[:4]]
y1, x1, y2, x2, prob = *list(
map(lambda x: max(x, 0), (y1, x1, y2, x2))), bbox[-1]
crop_img = frame[y1:y2, x1:x2, ...]
img_shape = tuple(model.input.shape.as_list()[1:3])
img = cv2.resize(crop_img,
img_shape,
interpolation=cv2.INTER_AREA)
img = img[np.newaxis, ...]
img = img / 255.
y_pred = model.predict(img)
model_predict_bbox[bbox_idx].append({
'O': y_pred[0][0],
'X': y_pred[0][1],
'=': y_pred[0][2],
'A': y_pred[0][3]
})
else:
logger.exception(
'Read frame faile at index {}'.format(frame_idx))
y_preds[model_path].append([frame_idx, model_predict_bbox])
logger.info('Complete model predict - {}'.format(model_path))
# ensemble result
reference_result = list(y_preds.keys())[0]
reference_result = y_preds[reference_result]
savepath = Path(args.input)
savepath = savepath.parent / '{}_ensemble.txt'.format(
str(Path(args.video).stem))
with open(str(savepath), 'w') as f:
for idx in tqdm(range(len(reference_result))):
frame_idx, bboxes = reference_result[idx]
merge_bboxes = []
for bbox_idx, bbox in enumerate(bboxes):
y_pred = np.array([ \
[v[idx][-1][bbox_idx][-1][k] for v in y_preds.values()] \
for k in bbox[-1].keys()])
y_pred = np.average(y_pred, axis=-1)
merge_bboxes.append([*bbox[:-1], \
{k: y_pred[idx] for idx, k in enumerate(bbox[-1].keys())}])
f.write(str([frame_idx, merge_bboxes]))
if idx != len(reference_result) - 1:
f.write('\n')
logger.info('Save final result at {}'.format(str(savepath)))
cap.release()
def main(args):
logger = logging.getLogger(__name__)
log_handler(logger)
logger.info(args)
# setting GPU
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpus
num_gpus = len(args.gpus.split(','))
available_gpus = get_available_gpus()
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.5
K.tensorflow_backend.set_session(tf.Session(config=config))
logger.info('available_gpus: {}'.format(available_gpus))
assert len(args.models) == num_gpus
# prepare detection results
with open(args.input, 'r') as f:
bbox_results = f.readlines()
bbox_results = [eval(b.rstrip('\n')) for b in bbox_results]
# prepare model predict results in multithread
queue = Queue()
videos = [args.video] * num_gpus
detect = [bbox_results] * num_gpus
thread_jobs = []
for model_args in zip(available_gpus, videos, args.models, detect):
model_args = list(model_args)
model_args.append(queue)
job = Thread(target=inference, args=model_args)
thread_jobs.append(job)
job.start()
for j in thread_jobs:
logger.info('join jobs {}'.format(j))
j.join()
logger.info('Complete multithread inference')
logger.info('Current thread {}'.format(current_thread()))
y_preds = {}
while not queue.empty():
model_path, y_pred = queue.get()
y_preds[model_path] = y_pred
# ensemble result
logger.info(y_preds.keys())
reference_result = list(y_preds.keys())[0]
reference_result = y_preds[reference_result]
savepath = Path(args.input)
savepath = savepath.parent / '{}_ensemble.txt'.format(
str(Path(args.video).stem))
with open(str(savepath), 'w') as f:
for idx in tqdm(range(len(reference_result))):
frame_idx, bboxes = reference_result[idx]
merge_bboxes = []
for bbox_idx, bbox in enumerate(bboxes):
y_pred = np.array([ \
[v[idx][-1][bbox_idx][-1][k] for v in y_preds.values()] \
for k in bbox[-1].keys()])
y_pred = np.average(y_pred, axis=-1)
merge_bboxes.append([*bbox[:-1], \
{k: y_pred[idx] for idx, k in enumerate(bbox[-1].keys())}])
f.write(str([frame_idx, merge_bboxes]))
if idx != len(reference_result) - 1:
f.write('\n')
logger.info('Save final result at {}'.format(str(savepath)))