NxFizzBuzz
FizzBuzz problem got solved by deep learning!
This repository shows an inductive programming demo using Nx, a multi-dimensional tensors library for Elixir.
Usage
# generate a dataset
{features, labels} = NxFizzBuzz.Dataset.generate_dataset(10000, 20)
# train the model by the dataset
params = NxFizzBuzz.Model.fit(features, labels, epoch: 100, batch_size: 50, hidden_size: 8)
# predict answers
1..100
|> Enum.each(fn n ->
NxFizzBuzz.predict_fizz_buzz(params, n)
|> IO.puts()
end)Evaluation
It took about 500 secs to train the model using 10,000 data in 100 times epochs. Then 100% of accuracy got achieved.
For the experiment, I used MacBook Pro (2018) with 2.7 GHz Quad Cores Intel Core i7. EXLA was not enabled and no GPUs were used.
$ time mix run examples/fizz_buzz.exs
(snip)
97: prediction: 97, answer: 97, matched?: true
98: prediction: 98, answer: 98, matched?: true
99: prediction: Fizz, answer: Fizz, matched?: true
100: prediction: Buzz, answer: Buzz, matched?: true
================
Accuracy: 1.0
________________________________________________________
Executed in 487.17 secs fish external
usr time 477.32 secs 126.00 micros 477.32 secs
sys time 48.84 secs 688.00 micros 48.84 secs
The result shown above is quite obvious because the dataset genearted by genuin FizzBuzz function are totally explanable deterministically. You have to notice it's a toy trial for an Nx demo ;)
Acknowledgement
This code is heavily inspired by the article below:
Author
Kentaro Kuribayashi <kentarok@gmail.com>