def _rollback(self, is_final=False):
# TODO: update rollback mechanism
if not self._allow_rollback and not is_final:
return
# trigger event for callbacks
self._callback.event(TrainSignal.ROLLBACK)
# default rollback procedure
if self._save_path is not None and os.path.exists(self._save_path):
self._show_noti("[%s] Rollback from: %s" %
(ctext('MainLoop', 'red'), self._save_path))
# restore previous checkpoint immediately
N.deserialize(self._save_path, force_restore_vars=True)
# otherwise, load stored variables from RAM
elif self._best_object is not None:
self._show_noti("[%s] Rollback to the best stored object from RAM" %
(ctext('MainLoop', 'red')))
N.deserialize(path_or_data=self._best_object,
force_restore_vars=True)
def _rollback(self, is_final=False):
# TODO: update rollback mechanism
if not self._allow_rollback and not is_final:
return
# trigger event for callbacks
self._callback.event(TrainSignal.ROLLBACK)
# default rollback procedure
if self._save_path is not None and os.path.exists(self._save_path):
self._show_noti("[%s] Rollback from: %s" %
(ctext('MainLoop', 'red'), self._save_path))
# restore previous checkpoint immediately
N.deserialize(self._save_path, force_restore_vars=True)
# otherwise, load stored variables from RAM
elif self._best_object is not None:
self._show_noti("[%s] Rollback to the best stored object from RAM" %
(ctext('MainLoop', 'red')))
N.deserialize(path_or_data=self._best_object,
force_restore_vars=True)
n_speakers = len(all_speakers) + 1
# ===========================================================================
# Create the network
# ===========================================================================
inputs = [
K.placeholder(shape=(None, ) + shape[1:],
dtype='float32',
name='input%d' % i)
for i, shape in enumerate(as_tuple_of_shape(train.shape))
]
X = inputs[0]
y = inputs[1]
print("Inputs:", ctext(inputs, 'cyan'))
# ====== the network ====== #
if os.path.exists(MODEL_PATH):
x_vec = N.deserialize(path=MODEL_PATH, force_restore_vars=True)
else:
TRAIN_MODEL = True
with N.args_scope(
['TimeDelayedConv',
dict(time_pool='none', activation=K.relu)],
['Dense', dict(activation=K.linear, b_init=None)],
['BatchNorm', dict(activation=K.relu)]):
x_vec = N.Sequence([
N.Dropout(level=0.3),
N.TimeDelayedConv(n_new_features=512, n_time_context=5),
N.TimeDelayedConv(n_new_features=512, n_time_context=5),
N.TimeDelayedConv(n_new_features=512, n_time_context=7),
N.Dense(512),
N.BatchNorm(),
N.Dense(1500),
all_speakers) = prepare_dnn_data(
recipe=args.recipe, feat=FEAT, utt_length=args.l)
n_speakers = len(all_speakers) + 1
# ===========================================================================
# Create the network
# ===========================================================================
inputs = [K.placeholder(shape=(None,) + shape[1:],
dtype='float32',
name='input%d' % i)
for i, shape in enumerate(as_tuple_of_shape(train.shape))]
X = inputs[0]
y = inputs[1]
print("Inputs:", ctext(inputs, 'cyan'))
# ====== the network ====== #
if os.path.exists(MODEL_PATH):
x_vec = N.deserialize(path=MODEL_PATH, force_restore_vars=True)
else:
TRAIN_MODEL = True
with N.args_scope(
['TimeDelayedConv', dict(time_pool='none', activation=K.relu)],
['Dense', dict(activation=K.linear, b_init=None)],
['BatchNorm', dict(activation=K.relu)]
):
x_vec = N.Sequence([
N.Dropout(level=0.3),
N.TimeDelayedConv(n_new_features=512, n_time_context=5),
N.TimeDelayedConv(n_new_features=512, n_time_context=5),
N.TimeDelayedConv(n_new_features=512, n_time_context=7),
N.Dense(512), N.BatchNorm(),
# Create the network
# ===========================================================================
inputs = [K.placeholder(shape=(None,) + shape[1:],
dtype='float32',
name='input%d' % i)
for i, shape in enumerate(as_tuple_of_shape(train.shape))]
X = inputs[0]
y = inputs[1]
print("Inputs:", ctext(inputs, 'cyan'))
# ====== get all saved model if exists ====== #
all_models = [os.path.join(EXP_DIR, i) for i in os.listdir(EXP_DIR)]
all_models = sorted([i for i in all_models if MODEL_PATH + '.' in i])
# ====== the network ====== #
if len(all_models) > 0:
print("Load model at:", ctext(all_models[SCORE_SYSTEM_ID], 'cyan'))
x_vec = N.deserialize(path=all_models[SCORE_SYSTEM_ID],
force_restore_vars=True)
else:
with N.args_scope(
['TimeDelayedConv', dict(time_pool='none', activation=K.relu)],
['Dense', dict(activation=K.linear, b_init=None)],
['BatchNorm', dict(activation=K.relu)]
):
x_vec = N.Sequence([
N.Dropout(level=0.3),
N.TimeDelayedConv(n_new_features=512, n_time_context=5),
N.TimeDelayedConv(n_new_features=512, n_time_context=5),
N.TimeDelayedConv(n_new_features=512, n_time_context=7),
N.Dense(512), N.BatchNorm(),
N.Dense(1500), N.BatchNorm(),
# ===========================================================================
# mapping from
# dataset_name -> 'name': 1-D array [n_samples],
# # (path to original audio)
# 'path': 1-D array [n_samples],
# # Extracted latent vectors
# 'X': 2-D array [n_samples, n_latent_dim]}
# # speaker label or meta-data (e.g. 'test', 'enroll', 'unlabeled')
# 'y': 1-D array [n_samples],
all_vectors = {}
# ===========================================================================
# Extract the x-vector for enroll and trials
# ===========================================================================
if 'xvec' == SCORE_SYSTEM_NAME:
# ====== load the network ====== #
x_vec = N.deserialize(path=final_sys, force_restore_vars=True)
# ====== get output tensors ====== #
y_logit = x_vec()
y_proba = tf.nn.softmax(y_logit)
X = K.ComputationGraph(y_proba).placeholders[0]
z = K.ComputationGraph(y_proba).get(roles=N.Dense,
scope='LatentOutput',
beginning_scope=False)[0]
f_z = K.function(inputs=X, outputs=z, training=False)
print('Inputs:', ctext(X, 'cyan'))
print('Latent:', ctext(z, 'cyan'))
# ====== recipe for feeder ====== #
recipe = prepare_dnn_feeder_recipe()
# ==================== extract x-vector from acoustic features ==================== #
for dsname, (ds_feat, ds_indices, ds_meta,
ds_path) in sorted(acoustic_features.items(),