Flux.jl/examples/char-rnn.jl

22 lines
590 B
Julia
Raw Normal View History

2016-10-28 23:16:39 +00:00
using Flux
2016-10-29 22:36:39 +00:00
getseqs(chars, alphabet) = sequences((onehot(Float32, char, alphabet) for char in chars), 50)
getbatches(chars, alphabet) = batches((getseqs(part, alphabet) for part in chunk(chars, 50))...)
2016-10-28 23:16:39 +00:00
2016-10-29 22:36:39 +00:00
input = readstring("$(homedir())/Downloads/shakespeare_input.txt")
2016-10-30 14:12:32 +00:00
alphabet = unique(input)
N = length(alphabet)
2016-10-28 23:16:39 +00:00
2016-10-30 14:12:32 +00:00
Xs, Ys = getbatches(input, alphabet), getbatches(input[2:end], alphabet)
2016-10-28 23:16:39 +00:00
model = Chain(
2016-10-30 14:12:32 +00:00
Input(N),
Recurrent(N, 128, N),
2016-10-28 23:16:39 +00:00
softmax)
2016-10-29 22:36:39 +00:00
m = tf(unroll(model, 50))
2016-10-28 23:16:39 +00:00
2016-10-30 14:12:32 +00:00
Flux.train!(m, Xs, Ys, η = 0.2e-3, epoch = 1)
2016-10-28 23:16:39 +00:00
2016-10-30 14:12:32 +00:00
map(c->onecold(c, alphabet), m(first(first(first(train)))))