n0e1 / lstm-poetry-word-based Goto Github PK

Similar to https://github.com/dvictor/lstm-poetry but this time, instead of characters, we represent entire words as (embedding) vectors.

TensorFlow version 0.7.1

You can download the trained NN from https://yadi.sk/d/Q0W0v-6yqxGW6
This is a rather small model, for better results you can train a larger net by adjusting the configuration parameters in train.py

• train.py train your NN using input.txt from your WORK_DIR
• generate.py generate text

Change WORK_DIR in each file to specify your work directory.

Input text:

green people floating
the morning has

Output:

green people floating
the morning has just begun

that's what the world has left to do
if you don't care how it's gotta be

i'd like to see you laughing at me

i can hear you say,

i can see you in the eyes of a smile
i'll be standing by your side

i can't stop the tears
with it black
i can see the white lines

can i even get your love?

Occasional <unk> in the output are caused by the limit imposed to the vocabulary size.
The vocabulary is composed of vocab_size most used words in the training set.
The words in the input that are not part of the vocabulary are encoded as <unk> so the NN will also generate some when it "feels like".