Install to use tts
pip install pyttsx3
This version of kears works best for me:
pip install --user keras==2.3.1
Notes
- large batch size is better, training can take longer to converge
- high dropout (>0.5) can help if dataset is small
- validation wastes a lot of time, don't do it unless you have a reason
Add some different losses
# line 200 in textgenrnn.py
base_lr = 4e-3
# scheduler function must be defined inline.
def lr_linear_decay(epoch):
return (base_lr * (1 - (epoch / num_epochs)))
final_decay = 0.04
def exp_decay(epoch):
lr = (base_lr * (final_decay ** ( epoch / num_epochs ) ) )
print("LR : {:.6f}".format(lr))
return lr
step_size = 14
from math import sin, pi, cos
def sin_decay(epoch):
decay = base_lr*(final_decay**(epoch/num_epochs))
inner_sin = (2*pi*(epoch%step_size))/step_size
lr = decay + (0.5*(base_lr*final_decay*(sin(inner_sin)+1)))
print("LR : {:.6f}".format(lr))
return lr
def cos_decay(epoch):
lr = base_lr * cos( (pi*epoch) / (2*num_epochs))
print("LR : {:.6f}".format(lr))
return lr
Also change the lr in the optimizer
optimizer=Adam(lr=base_lr)