class Adapt_Pretrainer(Trainer): def __init__(self, **kwargs): super(Adapt_Pretrainer, self).__init__(trainer_type='pretraining', **kwargs) def build(self): self.model = Adapt(**self.args) self.model.tensorboard_init() self.model.init_all()
class Pretrained_Inference(Trainer):
def __init__(self, separator, name, **kwargs):
super(Pretrained_Inference, self).__init__(trainer_type=name, **kwargs)
self.separator = separator
def build(self):
self.args.update({'pretraining':True})
self.model = Adapt(**self.args)
self.model.create_saver()
self.model.restore_model(self.args['model_folder'])
def build(self): self.model = Adapt.load(self.args['model_folder'], self.args) # Expanding the graph with enhance layer self.model.connect_front(self.separator) self.model.sepNet.output = self.model.sepNet.separate self.model.back self.model.create_saver() self.model.restore_model(self.args['model_folder']) self.model.finish_construction() self.model.initialize_non_init()
def build(self): self.model = Adapt.load(self.args['model_folder'], self.args) # Restoring the front layer: # Expanding the graph with enhance layer self.model.connect_front(self.separator) self.model.sepNet.output = self.model.sepNet.enhance self.model.back self.model.create_saver() self.model.restore_model(self.args['model_folder']) # Initialize only non restored values self.model.initialize_non_init()
def build(self):
if self.args['model_previous'] is not None:
self.model = Adapt.load(self.args['model_previous'], self.args)
self.model.connect_front(self.separator)
self.model.sepNet.output = self.model.sepNet.prediction
self.model.cost_model = self.model.sepNet.cost
self.model.back # To save the back values !
self.model.create_saver()
self.model.restore_model(self.args['model_previous'])
self.model.finish_construction()
self.model.freeze_all_with('front/')
self.model.freeze_all_with('back/')
self.model.optimize
self.model.tensorboard_init()
self.model.initialize_non_init()
else:
self.model = Adapt.load(self.args['model_folder'], self.args)
self.model.connect_only_front_to_separator(self.separator)
# Initialize only non restored values
self.model.initialize_non_init()
def build(self):
self.model = Adapt.load(self.args['model_folder'], self.args)
# Expanding the graph with enhance layer
self.model.connect_front(self.separator)
self.model.sepNet.output = self.model.sepNet.separate
self.model.back
self.model.create_saver()
self.model.restore_model(self.args['model_folder'])
self.model.cost_model = self.model.cost
self.model.finish_construction()
self.model.freeze_all_except('prediction', 'speaker_centroids')
self.model.optimize
self.model.tensorboard_init()
# Initialize only non restored values
self.model.initialize_non_init()
def build_model(self):
self.model = Adapt.load(self.args['model_folder'], self.args)
# Restoring previous Model:
self.model.restore_front_separator(self.args['model_folder'],
self.separator)
# Expanding the graph with enhance layer
with self.model.graph.as_default():
self.model.sepNet.output = self.model.sepNet.enhance
self.model.cost_model = self.model.sepNet.enhance_cost
self.model.finish_construction()
self.model.freeze_all_except('enhance')
self.model.optimize
self.model.tensorboard_init()
# Initialize only non restored values
self.model.initialize_non_init()
def build_model(self):
self.model = Adapt.load(self.args['model_folder'], self.args)
# Expanding the graph with enhance layer
with self.model.graph.as_default():
self.model.connect_front(self.separator)
self.model.sepNet.output = self.model.sepNet.separate
self.model.back
self.model.restore_model(self.args['model_folder'])
self.model.cost_model = self.model.cost
self.model.finish_construction()
self.model.optimize
self.model.tensorboard_init()
# Initialize only non restored values
self.model.initialize_non_init()
def build(self):
self.model = Adapt.load(self.args['model_folder'], self.args)
self.model.front
self.model.pretraining = True
self.model.separator
self.model.back
self.model.create_saver()
self.model.restore_model(self.args['model_folder'])
self.model.enhance
self.model.cost_model = self.model.enhance_cost
self.model.finish_construction()
self.model.freeze_all_with('front/')
self.model.freeze_all_with('back/')
self.model.optimize
self.model.tensorboard_init()
self.model.initialize_non_init()
def build(self):
self.model = Adapt.load(self.args['model_folder'], self.args)
# Restoring the front layer:
# Expanding the graph with enhance layer
self.model.connect_front(self.separator)
self.model.sepNet.output = self.model.sepNet.enhance
self.model.back
self.model.create_saver()
self.model.restore_model(self.args['model_folder'])
self.model.cost_model = self.model.cost_finetuning
self.model.finish_construction()
to_train = []
for var in self.model.trainable_variables:
for p in self.args['train']:
if p in var.name:
to_train.append(var)
self.model.trainable_variables = to_train
# self.model.freeze_all_except('prediction', 'speaker_centroids', 'enhance')
self.model.optimize
self.model.tensorboard_init()
# Initialize only non restored values
self.model.initialize_non_init()
males = H5PY_RW()
males.open_h5_dataset('test_raw.h5py', subset=males_keys(H5_dico))
males.set_chunk(5 * 4 * 512)
males.shuffle()
print 'Male voices loaded: ', males.length(), ' items'
fem = H5PY_RW()
fem.open_h5_dataset('test_raw.h5py', subset=females_keys(H5_dico))
fem.set_chunk(5 * 4 * 512)
fem.shuffle()
print 'Female voices loaded: ', fem.length(), ' items'
Mixer = Mixer([males, fem], with_mask=False, with_inputs=True)
adapt_model = Adapt.load('jolly-firefly-9628',
pretraining=False,
separator=DPCL)
# adapt_model.init()
print 'Model DAS created'
testVar = raw_input("Model loaded : Press Enter")
cost_valid_min = 1e10
Mixer.select_split(0)
learning_rate = 0.01
for i in range(config.max_iterations):
X_in, X_mix, Ind = Mixer.get_batch(1)
if (i + 1) % 100 == 0:
learning_rate /= 10
c = adapt_model.train(X_mix, X_in, learning_rate, i)
def build(self): self.model = Adapt.load(self.args['model_folder'], self.args) # Restoring previous Model: self.model.connect_enhance_to_separator(self.separator) self.model.initialize_non_init()
def build(self): self.model = Adapt.load(self.args['model_folder'], self.args) self.model.connect_only_front_to_separator(self.separator) # Initialize only non restored values self.model.initialize_non_init()
def build(self): self.model = Adapt(**self.args) self.model.tensorboard_init() self.model.init_all()
config_model["smooth_size"] = 10
config_model["alpha"] = learning_rate
config_model["reg"] = 1e-3
config_model["beta"] = 0.1
config_model["rho"] = 0.01
config_model["same_filter"] = True
config_model["optimizer"] = 'Adam'
####
####
adapt_model = Adapt(config_model=config_model,
pretraining=True,
folder='pretraining')
adapt_model.tensorboard_init()
adapt_model.init()
print 'Total name :'
print adapt_model.runID
# nb_iterations = 500
mixed_data.adjust_split_size_to_batchsize(batch_size)
nb_batches = mixed_data.nb_batches(batch_size)
nb_epochs = 2
time_spent = [0 for _ in range(5)]
for epoch in range(nb_epochs):
config_model["beta"] = 0.1
config_model["rho"] = 0.01
idd = ''.join('-{}={}-'.format(key, val)
for key, val in sorted(config_model.items()))
batch_size = 4
config_model["batch_size"] = batch_size
config_model["type"] = "Dense_train"
from models.adapt import Adapt
import config
full_id = 'soft-base-9900' + idd
folder = 'Dense_train'
model = Adapt(config_model=config_model, pretraining=False)
model.create_saver()
path = os.path.join(config.workdir, 'floydhub_model', "pretraining")
# path = os.path.join(config.log_dir, "pretraining")
model.restore_model(path, full_id)
## Connect DAS model to the front end
from models.dense import Dense_net as Dense
with model.graph.as_default():
model.connect_front(Dense)
model.sepNet.output = model.sepNet.prediction
model.back
model.cost
full_id = 'cold-dust-9076' + idd
path = os.path.join(config.model_root if not config.floydhub else '/model2',
'log', config_model["type"])
####
#### NEW MODEL CONFIGURATION
####
config_model["type"] = "L41_finetuning"
learning_rate = 0.001
batch_size = 64
config_model["chunk_size"] = chunk_size
config_model["alpha"] = learning_rate
config_model["batch_size"] = batch_size
model = Adapt(config_model=config_model, pretraining=False)
with model.graph.as_default():
model.connect_front(L41Model)
var_list = [v for v in tf.global_variables() if ('front' in v.name)]
model.create_saver(var_list)
model.restore_model(path_adapt, full_id_adapt)
model.sepNet.prediction
model.sepNet.separate
model.sepNet.output = model.sepNet.enhance
var_list = [
v for v in tf.global_variables()
if ('prediction' in v.name or 'speaker_centroids' in v.name
or 'enhance' in v.name)
]
model.create_saver(var_list)
full_id = "frosty-fire-4612" + idd #### #### NEW MODEL CONFIG #### config_model["type"] = "L41_enhance" learning_rate = 0.001 batch_size = 8 config_model["chunk_size"] = chunk_size config_model["batch_size"] = batch_size config_model["alpha"] = learning_rate config_model["optimizer"] = 'Adam' config_model["reg"] = 1e-3 model = Adapt(config_model=config_model, pretraining=False) # Small modification for enhance #TODO with model.graph.as_default(): model.connect_front(L41Model) model.sepNet.output = model.sepNet.prediction model.create_saver() model.restore_model(path, full_id) model.sepNet.separate model.sepNet.output = model.sepNet.enhance model.cost = model.sepNet.enhance_cost model.freeze_variables() model.optimize model.tensorboard_init() init = model.non_initialized_variables()
males = H5PY_RW()
males.open_h5_dataset('test_raw.h5py', subset=males_keys(H5_dico))
males.set_chunk(5 * 4 * 512)
males.shuffle()
print 'Male voices loaded: ', males.length(), ' items'
fem = H5PY_RW()
fem.open_h5_dataset('test_raw.h5py', subset=females_keys(H5_dico))
fem.set_chunk(5 * 4 * 512)
fem.shuffle()
print 'Female voices loaded: ', fem.length(), ' items'
Mixer = Mixer([males, fem], with_mask=False, with_inputs=True)
adapt_model = Adapt()
print 'Model DAS created'
adapt_model.init()
cost_valid_min = 1e10
Mixer.select_split(0)
learning_rate = 0.005
for i in range(config.max_iterations):
X_in, X_mix, Ind = Mixer.get_batch(1)
c = adapt_model.train(X_mix, X_in, learning_rate, i)
print 'Step #', i, ' ', c
if i % 20 == 0: #cost_valid < cost_valid_min:
print 'DAS model saved at iteration number ', i, ' with cost = ', c
adapt_model.save(i)
config_model["beta"] = 0.1
config_model["rho"] = 0.01
idd = ''.join('-{}={}-'.format(key, val)
for key, val in sorted(config_model.items()))
batch_size = 1
config_model["batch_size"] = batch_size
config_model["type"] = "DAS_train_front"
from models.adapt import Adapt
import config
full_id = 'soft-base-9900' + idd
folder = 'DAS_train_front'
model = Adapt(config_model=config_model, pretraining=False)
model.create_saver()
path = os.path.join(config.workdir, 'floydhub_model', "pretraining")
# path = os.path.join(config.log_dir, "pretraining")
model.restore_model(path, full_id)
from models.das import DAS
model.connect_only_front_to_separator(DAS)
init = model.non_initialized_variables()
# Model creation
# Pretraining the model
for key, val in sorted(config_model.items())) full_id = "noisy-breeze-3898" + idd path = os.path.join(config.model_root, 'log', 'pretraining') #### #### NEW MODEL #### config_model["type"] = "L41_train_front" learning_rate = 0.01 batch_size = 8 config_model["chunk_size"] = 512 * 40 config_model["batch_size"] = batch_size config_model["alpha"] = learning_rate model = Adapt(config_model=config_model, pretraining=False) model.create_saver() model.restore_model(path, full_id) model.connect_only_front_to_separator(L41Model) init = model.non_initialized_variables() model.sess.run(init) print 'Total name :' print model.runID # nb_iterations = 500 mixed_data.adjust_split_size_to_batchsize(batch_size) nb_batches = mixed_data.nb_batches(batch_size) nb_epochs = 40
def build(self):
self.args.update({'pretraining':True})
self.model = Adapt(**self.args)
self.model.create_saver()
self.model.restore_model(self.args['model_folder'])
####
#### NEW MODEL CONFIGURATION
####
config_model["type"] = "DPCL_finetuning"
learning_rate = 0.001
batch_size = 2
config_model["chunk_size"] = chunk_size
config_model["alpha"] = learning_rate
config_model["batch_size"] = batch_size
folder = 'DPCL_finetuning'
model = Adapt(config_model=config_model, pretraining=False)
model.create_saver()
path = os.path.join(config.model_root, 'log', 'DPCL_train_front')
model.restore_model(path, full_id)
model.connect_front_back_to_separator(DPCL)
with model.graph.as_default():
model.create_saver()
model.restore_model(path, full_id)
# model.freeze_front()
model.optimize
model.tensorboard_init()
init = model.non_initialized_variables()
#### LOAD PREVIOUS MODEL
####
idd = ''.join('-{}={}-'.format(key, val)
for key, val in sorted(config_model.items()))
config_model["type"] = "DPCL_train_front"
learning_rate = 0.01
batch_size = 32
config_model["batch_size"] = batch_size
config_model["alpha"] = learning_rate
full_id = "long-term-4925" + idd
#full_id = 'jolly-sound-3162'+idd
folder = 'DPCL_train_front'
model = Adapt(config_model=config_model, pretraining=False)
model.create_saver()
path = os.path.join(config.model_root, 'log', 'pretraining')
model.restore_model(path, full_id)
model.connect_only_front_to_separator(DPCL)
init = model.non_initialized_variables()
model.sess.run(init)
print 'Total name :'
print model.runID
# nb_iterations = 500
mixed_data.adjust_split_size_to_batchsize(batch_size)
