makemore

makemore takes one text file as input, where each line is assumed to be one training thing, and generates more things like it. Under the hood, it is an autoregressive character-level language model, with a wide choice of models from bigrams all the way to a Transformer (exactly as seen in GPT). For example, we can feed it a database of names, and makemore will generate cool baby name ideas that all sound name-like, but are not already existing names. Or if we feed it a database of company names then we can generate new ideas for a name of a company. Or we can just feed it valid scrabble words and generate english-like babble.

This is not meant to be too heavyweight library with a billion switches and knobs. It is one hackable file, and is mostly intended for educational purposes. PyTorch is the only requirement.

Current language model neural nets implemented:

• Bigram (one character simply predicts a next one with a lookup table of counts)
• Bag of Words
• MLP, along the lines of Bengio et al. 2003
• RNN, along the lines of Sutskever et al. 2011
• GRU, following Kyunghyun Cho et al. 2014
• Transformer, following Vaswani et al. 2017

Usage

The included names.txt dataset, as an example, has the most common 32K names takes from ssa.gov for the year 2018. It looks like:

emma
olivia
ava
isabella
sophia
charlotte
...

Let's point the script at it:

$ python makemore.py -i names.txt -o names

Training progress and logs and model will all be saved to the working directory names. The default model is a super tiny 200K param transformer; Many more training configurations are available - see the argparse and read the code. Training does not require any special hardware, it runs on my Macbook Air and will run on anything else, but if you have a GPU then training will fly faster. As training progresses the script will print some samples throughout. However, if you'd like to sample manually, you can use the --sample-only flag, e.g. in a separate terminal do:

$ python makemore.py -i names.txt -o names --sample-only

This will load the best model so far and print more samples on demand. Here are some unique baby names that get eventually generated from current default settings (test logprob of ~1.92, though much lower logprobs are achievable with some hyperparameter tuning):

dontell
khylum
camatena
aeriline
najlah
sherrith
ryel
irmi
taislee
mortaz
akarli
maxfelynn
biolett
zendy
laisa
halliliana
goralynn
brodynn
romima
chiyomin
loghlyn
melichae
mahmed
irot
helicha
besdy
ebokun
lucianno

Have fun!

License

MIT