def get_coco(root, image_set, transforms, mode='instances'):
anno_file_template = "{}_{}2017.json"
PATHS = {
"train": ("train2017",
os.path.join("annotations",
anno_file_template.format(mode, "train"))),
"val": ("val2017",
os.path.join("annotations",
anno_file_template.format(mode, "val"))),
# "train": ("val2017", os.path.join("annotations", anno_file_template.format(mode, "val")))
}
t = [ConvertCocoPolysToMask()]
if transforms is not None:
t.append(transforms)
transforms = T.Compose(t)
img_folder, ann_file = PATHS[image_set]
img_folder = os.path.join(root, img_folder)
ann_file = os.path.join(root, ann_file)
dataset = CocoDetection(img_folder, ann_file, transforms=transforms)
if image_set == "train":
dataset = _coco_remove_images_without_annotations(dataset)
# dataset = torch.utils.data.Subset(dataset, [i for i in range(500)])
return dataset
def BOT_recognition(model_BOT, img):
transform = T.Compose([
T.Resize((256, 128)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
loader = DataLoader(
Load_person(img, transform=transform),
batch_size=1,
shuffle=False,
num_workers=0,
pin_memory=True,
drop_last=False,
)
model_BOT.eval()
with torch.no_grad():
for batch_idx, img2 in enumerate(loader):
if torch.cuda.is_available(): img2 = img2.cuda()
# score = model_BOT(img2)
gender_outputs, staff_outputs, customer_outputs, stand_outputs, sit_outputs, phone_outputs = model_BOT(
img2)
staff_con = float(Confidence(staff_outputs, 1))
customer_con = float(Confidence(customer_outputs, 1))
stand_con = float(Confidence(stand_outputs, 1))
sit_con = float(Confidence(sit_outputs, 1))
play_with_phone_con = float(Confidence(phone_outputs, 1))
male_con = float(Confidence(gender_outputs, 0))
female_con = float(Confidence(gender_outputs, 1))
staff_con = 1 if staff_con > 0.5 else 0
customer_con = 1 if customer_con > 0.5 else 0
stand_con = 1 if stand_con > 0.5 else 0
sit_con = 1 if sit_con > 0.5 else 0
play_with_phone_con = 1 if play_with_phone_con > 0.1 else 0
male_con = 1 if male_con > 0.5 else 0
female_con = 1 if female_con > 0.5 else 0
return staff_con, customer_con, stand_con, sit_con, play_with_phone_con, male_con, female_con
def get_transform(args):
melkwargs = {
'n_mels': args.nfilt,
'n_fft': args.nfft,
'win_length': int(args.winlen * args.signal_sr),
'hop_length': int(args.winstep * args.signal_sr),
}
args.signal_samples = int(args.signal_sr * args.signal_len)
args.signal_width = int(
np.ceil((args.signal_len - args.winlen) / args.winstep) + 1)
if args.features_name.lower() == 'logfbes':
features = transforms.Compose([
transforms.LogFBEs(args.signal_sr, args.winlen, args.winstep,
args.nfilt, args.nfft, args.preemph),
transforms.ToTensor(),
])
args.nfeature = args.nfilt
args.signal_width = int(
np.ceil((args.signal_len - args.winlen) / args.winstep) + 1)
elif args.features_name.lower() == 'mfccs':
features = transforms.Compose([
transforms.MFCCs(args.signal_sr, args.winlen, args.winstep,
args.numcep, args.nfilt, args.nfft, args.preemph,
args.ceplifter),
transforms.ToTensor(),
])
args.nfeature = args.numcep
args.signal_width = int(
np.ceil((args.signal_len - args.winlen) / args.winstep) + 1)
elif args.features_name.lower() == 'ta.mfccs':
features = transforms.Compose([
# transforms.ToTensor(),
torchaudio.transforms.MFCC(sample_rate=args.signal_sr,
n_mfcc=args.numcep,
melkwargs=melkwargs),
])
args.nfeature = args.numcep
args.signal_width = 1 + args.signal_samples // melkwargs['hop_length']
# args.signal_width = 81;
elif args.features_name.lower() == 'ta.logfbes':
log_offset = 1e-6
features = transforms.Compose([
# transforms.ToTensor(),
torchaudio.transforms.MelSpectrogram(sample_rate=args.signal_sr,
**melkwargs),
transforms.Lambda(lambda t: torch.log(t + log_offset)),
])
args.nfeature = args.nfilt
args.signal_width = 1 + args.signal_samples // melkwargs['hop_length']
# args.signal_width = 81;
else:
raise Exception(
'--features_name should be one of {LogFBEs | MFCCs | ta.MFCCs | ta.LogFBEs}'
)
# Composing transforms
test_trasform = transforms.Compose([
features,
transforms.Lambda(lambda x: x.unsqueeze(0)),
])
silence_transform = transforms.Compose([
transforms.Lambda(lambda x: x * random.uniform(0.0, args.silence_vol)),
test_trasform,
])
args.signal_samples = int(args.signal_sr * args.signal_len)
args.bkg_noise_path = 'background_noise'
train_transform = transforms.Compose([
# transforms.TimeShifting2(shift_min=args.shift_min, shift_max=args.shift_max),
# transforms.RandomApplyTransform(p=args.p_transform, transform=transforms.AddNoise3(
# os.path.join(args.data_root, args.bkg_noise_path),
# args.noise_vol, args.signal_samples, args.signal_sr)),
test_trasform,
# torchaudio.transforms.TimeMasking(args.mask_time),
# torchaudio.transforms.TimeMasking(args.mask_time),
# torchaudio.transforms.FrequencyMasking(args.mask_freq),
])
return {
'train': train_transform,
'val': test_trasform,
'test': test_trasform
}, silence_transform
def get_transform(train):
transforms = []
transforms.append(T.ToTensor())
if train:
transforms.append(T.RandomHorizontalFlip(0.5))
return T.Compose(transforms)
def get_transform(args):
melkwargs = {
'n_mels': args.nfilt,
'n_fft': args.nfft,
'win_length': int(args.winlen * args.signal_sr),
'hop_length': int(args.winstep * args.signal_sr),
}
args.signal_samples = int(args.signal_sr * args.signal_len)
args.signal_width = 1 + args.signal_samples // melkwargs['hop_length']
args.signal_samples = int(args.signal_sr * args.signal_len)
args.bkg_noise_path = 'background_noise'
if args.features_name.lower() == 'mfccs':
features = transforms.Compose([
torchaudio.transforms.MFCC(sample_rate=args.signal_sr, n_mfcc=args.numcep, log_mels=True, melkwargs=melkwargs),
transforms.Lambda(lambda x: x[1:]),
])
args.nfeature = args.numcep - 1
elif args.features_name.lower() == 'logfbes':
log_offset = 1e-6
features = transforms.Compose([
torchaudio.transforms.MelSpectrogram(sample_rate=args.signal_sr, **melkwargs),
transforms.Lambda(lambda t: torch.log(t + log_offset)),
])
args.nfeature = args.nfilt
else:
raise Exception('--features_name should be one of {LogFBEs | MFCCs}')
def debug(x):
print(x.size())
return x
# Composing transforms
test_trasform = transforms.Compose([
features,
transforms.Lambda(lambda x: x.unsqueeze(0)),
])
silence_transform = transforms.Compose([
transforms.Lambda(lambda x: x * random.uniform(0.0, args.silence_vol)),
test_trasform,
])
if args.no_augmentations:
train_transform = test_trasform
else:
train_transform = transforms.Compose([
transforms.TimeShifting2(shift_min=args.shift_min, shift_max=args.shift_max),
transforms.AddNoise3(
os.path.join(args.data_root, args.bkg_noise_path),
args.noise_vol, args.signal_samples, args.signal_sr),
test_trasform,
torchaudio.transforms.TimeMasking(args.mask_time),
torchaudio.transforms.TimeMasking(args.mask_time),
torchaudio.transforms.FrequencyMasking(args.mask_freq),
])
return {'train': train_transform, 'val': test_trasform, 'test': test_trasform}, silence_transform
handlers=[
logging.FileHandler(
os.path.join(
log_dir,
f'{datetime.now().strftime("%Y%m%d-%H%M%S")}_{args.model}.log')
),
logging.StreamHandler()
])
####### BASE TRANSFORMS ##########
base_transforms = T.Compose([
T.DataCreator(
), # convert value-dictionary to instance of MultiNodeData-class
T.DataLabeler(min_occurrence=0.5), # label nodes based on tag occurrences
T.NodeSelector(include_nodes=args.include_nodes
), # only use subset of nodes if required
T.MetricSelector(include_metrics=args.include_metrics
), # only use subset of metrics if required
T.AnomalySelector(exclude_anomalies=args.exclude_anomalies
), # filter anomalies if exclusions are specified
T.NodeConnector(
use_synthetic_data=use_synthetic_data) # add adj-matrix (sparse)
])
logging.info(f"Base transforms: {base_transforms}")
########## PREPARE DATA ##########
logging.info("Read dataset...")
data_root = args.data_root
dataset = MultiNodeDataset(data_root,
args.input_dimension,
args.stride,
use_synthetic_data=use_synthetic_data,
start_time=args.start_time,