def __init__(self, set_id=None, set_name=None, conf=conf_default):
self.connection = conn.get();
self.connection.register([DataSet])
if set_id == None:
self.train_set = self.connection.DataSet()
self.train_set.config = conf
self.train_set.name = set_name
self.train_set.save()
else:
self.train_set = self.connection.DataSet.one({"_id": ObjectId(set_id)})
self.folder_path = training_folder + '/' + self.generate_name()
self.validation_path = self.folder_path + '/validation/'
def train(tracks, training_audio_path, conf, set_name, conn_str):
logger.log('begin training sequence on %s ' % training_audio_path)
# set up all model parameters
num_components = conf['num_components']
em_epsilon = conf['em_epsilon']
em_iter = conf['em_iter']
cv_type = conf['cv_type']
audio_freq = conf['audio_freq']
mfcc_step_size = conf['mfcc_step_size']
mfcc_block_size = conf['mfcc_block_size']
frames_per_second = audio_freq / mfcc_step_size
audio_block_size = frames_per_second * conf['sample_length']
audio_step_size = frames_per_second * conf['sample_step_size']
# set up Yaafe
afp, engine = get_engine(audio_freq, mfcc_block_size, mfcc_step_size)
connection = conn.get()
connection.register([TrackData])
result = []
for label, filename in tracks:
logger.log('begin processing %s.' % filename)
afp.processFile(engine, training_audio_path + '/' + filename)
output = engine.readAllOutputs()['mfcc']
mfcc = output
num_samples = mfcc.shape[0]
track_gmms = []
track = connection.TrackData()
track.label = label
track.set = set_name
for index in range(0, (num_samples - audio_block_size), audio_step_size):
with warnings.catch_warnings(record=True) as w:
mfcc_data = mfcc[index:index + audio_block_size]
classifier = GMM(n_components = num_components, cvtype = cv_type)
classifier.fit(mfcc_data, thresh = em_epsilon, n_iter = em_iter)
means = classifier._get_means().tolist()
if cv_type == 'diag':
covars = [np.diag(diag).tolist() for diag in classifier._get_covars()]
else:
covars = classifier._get_covars().tolist()
weights = classifier._get_weights().tolist()
for weight in weights:
if math.isnan(weight):
logger.warn("Warning: invalid GMM entry, skipping at label: %s, index %s" % (label, str(index)))
break
else:
track_gmms.append([means, covars, weights])
track.data = track_gmms
track.save()
result.append(track._id)
return result
