Localitate Gen

Tool to generate Romanian sounding place names.

Sampling the model

python3.7 main.py sample --count 10
vladesti
furdului
albiu
ciobana
dobra
dornenu
ludosteni
carjmeschelmoceni
popesti
valea lun

You can also use completion, by giving a suffix to the sample command

main.py sample --prefix ma --count 5
macesti
magura roman
mandrusului
mazesti
malea mlan

Training

The model can be training using the train command. An optional plot option can be added to show the plot.

python3.7 main.py train --plot
Paramters: 26624
Training data size: 116433
Vocab size: 34
Block size: 4
Batch size: 64
Embedding size: 10
Layer one size: 350
Learning rate: 0.12
Learning rate decay after: 30000
Training iterations: 65000
Dev loss 3.393, train loss 3.391
Dev loss 2.069, train loss 2.071
...
Dev loss 1.690, train loss 1.486