def launch_servers(config_fn, which_set=['train', 'valid'], verbose=False):
""""""
global DEFAULT_PORT
config = load_config(config_fn)
data_servers = {}
ports = {}
port = DEFAULT_PORT
for target in config.target:
data_servers[target] = {}
ports[target] = {}
for dset in which_set:
if dset == 'valid':
ports[target][dset] = port + 1
else:
ports[target][dset] = port
args = [
'python', '-m', 'data_server.server', '--target', target,
'--which-set', dset, '--port',
str(ports[target][dset]), '--config-fn', config_fn
]
if verbose:
data_servers[target][dset] = subprocess.Popen(args)
else:
with open(os.devnull, 'w') as devnull:
data_servers[target][dset] = subprocess.Popen(
args, stdout=devnull, stderr=devnull)
port += 10
return data_servers, ports
def reload_config():
global c
global label_config_line
global analytics
c = load_config()
label_config_line = config_line_stripped(open(c['label_config']).read())
if analytics is None:
analytics = Analytics(label_config_line,
c.get('collect_analytics', True))
else:
analytics.update_info(label_config_line,
c.get('collect_analytics', True))
def initiate_data_server(target, which_set, port, config_fn):
"""
"""
config = load_config(config_fn)
print('Initialize Data Server...')
# dataset = MSD(
# target=target,
# which_set=which_set,
# config=config,
# )
dataset = MSDMel(target=target, which_set=which_set, config=config)
launch_data_server(dataset, port, config)
def __init__(self, model_fn, task, out_dir=None, hop_sz=1.):
""""""
super(MTLExtractor, self).__init__(task, out_dir, hop_sz, prob=True)
# load configuration for model
if os.path.exists(model_fn):
model_id = os.path.splitext(os.path.basename(model_fn))[0]
self.model_id = model_id.split('_state')[0]
model_state = joblib.load(model_fn)
self.config = namedtupled.map(model_state['config'])
else:
self.model_id = 'rnd'
# load default config and change task as rand
self.config = load_config('config/config.example.json')
self.config.target[0] = 'rand'
self.out_fn = os.path.join(
self.root, self.model_id + '_{}_feature.h5'.format(self.task))
self.targets = self.config.target
# load model
self.model = Model(self.config)
# variable set up
self.sr = self.config.hyper_parameters.sample_rate
self.length = self.config.hyper_parameters.patch_length
self.n_fft = self.config.hyper_parameters.n_fft
self.hop_sz_trn = self.config.hyper_parameters.hop_size
self.input = self.config.hyper_parameters.input
self.hop = int(self.hop_sz * self.sr)
sig_len = int(self.sr * self.length)
self.sig_len = sig_len - sig_len % self.hop_sz_trn
# prepare preprocessor if needed
if self.config.hyper_parameters.input == 'melspec':
self.melspec = MelSpectrogramGPU(2, self.sr, self.n_fft,
self.hop_sz_trn)
# set feature layer names
branch_at = self.config.hyper_parameters.branch_at
if isinstance(branch_at, (int, float)):
self.feature_layers = ['{}.fc'.format(t) for t in self.targets]
elif isinstance(branch_at, (str, unicode)) and branch_at == "fc":
self.feature_layers = ['fc']
self._prepare_db()
super(MTLExtractor, self).post_init()
self.hf.attrs['targets'] = [t.encode() for t in self.targets]
def main(config_fn, data_verbose=False):
""""""
# load config file
config_fn = os.path.abspath(config_fn)
config = load_config(config_fn)
# launch data servers
with data_context(config_fn,
which_set=['train', 'valid'],
verbose=data_verbose) as data_streams:
# initialize trainer
trainer = Trainer(config, data_streams)
# train
trainer.fit()
def data_context(config_fn, which_set, verbose=False):
""""""
config = load_config(config_fn)
# check remote
if hasattr(config.data_server, 'ports'):
ports = namedtupled.reduce(config.data_server.ports)
remote_server = True
else:
servers, ports = launch_servers(config_fn, which_set, verbose)
remote_server = False
streams = StreamManager(ports, config)
yield streams
# tear down data servers only when it's local
if not remote_server:
kill_servers(servers)
def test_model(config_fn, out_dir):
""""""
config = load_config(config_fn)
tasks = config.target
model = Model(config)
y, sr = load_test_audio(config)
x = np.repeat(y[None, None, 55040*2:55040*3], 2, axis=1)
X = np.array(
[np.abs(librosa.stft(y_, n_fft=1024, hop_length=256))
for y_ in x[0]])
np.save(os.path.join(out_dir, 'test_input.npy'), X)
Z = {}
for task in tasks:
Z[task] = model.predict(task, x)
np.save(
os.path.join(
out_dir, 'test_recon_{}.npy'.format(task)),
Z[task]
)
def reload_config():
global c
global label_config_line
global analytics
global ml_backend
global project
c = load_config()
label_config_line = config_line_stripped(open(c['label_config']).read())
if analytics is None:
analytics = Analytics(label_config_line,
c.get('collect_analytics', True))
else:
analytics.update_info(label_config_line,
c.get('collect_analytics', True))
# configure project
if project is None:
project = Project(label_config=label_config_line)
# configure machine learning backend
if ml_backend is None:
ml_backend_params = c.get('ml_backend')
if ml_backend_params:
ml_backend = MLBackend.from_params(ml_backend_params)
project.connect(ml_backend)
def test():
cfg = load_config('../utils/configurations/simulated_camera_24bit.json')
# luminance test
inputs = {'luminance': np.linspace(0.01, 10000, 100), # 0.01 to 10000 cd/m^2
'f_num': 8, 'exposure_time': 0.01, 'iso': 100}
unit_test(cfg, inputs)
# F-number test
inputs = {'f_num': np.logspace(1, 4, 10, base=2), # 2 to 16 with 0.3 stop
'luminance': 1000, 'exposure_time': 0.01, 'iso': 100}
unit_test(cfg, inputs)
# exposure time test
inputs = {'exposure_time': np.linspace(0.001, 0.5, 100), # 0.001 to 0.5 second
'luminance': 1000, 'f_num': 16, 'iso': 100}
unit_test(cfg, inputs)
# ISO speed test
inputs = {'iso': 100 * np.logspace(0, 4, 13, base=2), # ISO100 to ISO1600
'luminance': 1000, 'f_num': 8, 'exposure_time': 0.005}
unit_test(cfg, inputs)
plt.show(block=False)
import os
import tifffile
from simulation.simulated_camera import SimulatedCamera
from simulation.simulated_patterns import generate_dts_luminance, generate_dts_luminance_map
from cdp_calculator import CDPCalculator
from dts.roi_extractor import DTSRoIExtractor
from utils.misc import load_config
MAX_LUMINANCE = 50000 # luminance of the light source, in cd/m^2
NUM_IMAGES = 3 # capturing several frames will produce more accurate results
# load the configuration of the simulated camera
cfg = load_config('./utils/configurations/simulated_camera_24bit.json')
simulated_images_dir = './simulation/simulated_images' # directory to save the simulated images
os.makedirs(simulated_images_dir, exist_ok=True)
# generate simulated luminance map of CDP pattern in DTS device
dts_luminance = generate_dts_luminance(max_luminance=MAX_LUMINANCE)
dts_luminance_map = generate_dts_luminance_map(dts_luminance)
# a simulated camera that captures the CDP pattern
camera = SimulatedCamera(cfg)
for i in range(NUM_IMAGES):
print('Generating {}/{} HDR image'.format(i + 1, NUM_IMAGES))
image = camera.capture_hdr(dts_luminance_map, f_num=8, exposure_time=0.001, iso=125, frames=3, ev_step=4)
image = camera.tone_mapping(image)
save_path = os.path.join(simulated_images_dir, 'simulated_image_{}.tiff'.format(i))