def train(X,
y,
num_epochs,
train_fn,
val_fn=None,
train_iterator=None,
validation_iterator=None):
if train_iterator is None:
train_iterator = SpectTrainBatchIterator(100, 10, config)
if validation_iterator is None:
validation_iterator = SpectValidBatchIterator(100, 10, config)
if val_fn is not None:
train_splitter = sts.SpeakerTrainSplit(eval_size=0.25, sentences=8)
X_train, X_valid, y_train, y_valid = train_splitter(X, y)
else:
X_train = X
y_train = y
margin = 2
print("Start training")
confusion = analytics.ConfusionMatrix(len(set(y)))
theano.printing.pydotprint(train_fn,
outfile='/data/test.svg',
format='svg')
print('printed network')
for epoch in range(num_epochs):
start_time = time.time()
# Training
train_err = 0
train_batches = 0
for Xb, yb in train_iterator.iterate(X_train, y_train):
train_err += train_fn(Xb, yb)
train_batches += 1
# Validation
if val_fn is not None:
val_err = 0
val_batches = 0
for Xb, yb in validation_iterator.iterate(X_valid, y_valid):
err, class_predictions = val_fn(Xb, yb)
val_err += err
confusion.add_predictions(yb, class_predictions)
val_batches += 1
accs = confusion.calculate_accuracies()
confusion.mat.fill(0)
print("Epoch {} of {} took {:.3f}s".format(epoch + 1, num_epochs,
time.time() - start_time))
print(" training loss:\t\t{:.6f}".format(train_err / train_batches))
if val_fn is not None:
print(" validation loss:\t\t{:.6f}".format(val_err / val_batches))
print(" validation accuracy:\t\t{:.2f} %".format(accs.sum() /
len(accs) * 100))
sys.stdout.flush()
def create_train_data(self):
with open(get_speaker_pickle(self.training_data), 'rb') as f:
(X, y, speaker_names) = pickle.load(f)
splitter = sts.SpeakerTrainSplit(0.2, 10)
X_t, X_v, y_t, y_v = splitter(X, y)
return X_t, y_t, X_v, y_v
def _split_train_val_data(self, X, y):
splitter = sts.SpeakerTrainSplit(0.2)
X_t, X_v, y_t, y_v = splitter(X, y)
return X_t, y_t, X_v, y_v
