playwith_rnn

Hey there!

This repository describes an easy implementation and application of Recurrent Neural Network. I used keras as deep learning library "Using Tensorflow Backend" on CPU (yep!). I experimented with various types of text datasets as input to the model that then generates the probability distribution which is sampled to return the next best output for a given sequence of characters.

Names:

This directory mainly contains a generator.py file that trains on small LSTM model and generates new names. That simple. To execute it:

• Run generator.py.
• It will train the model now. It will take time depending on your CPU/GPU configuration.
• When done, check the generated names in temp_output.txt.

Text - character and Text - word:

The Text directory contains a main.py file that calls:

1. text_io.py file handles the input and output functions for LSTM model to generate some more "similar" text. It reads the text file, then splits the content into a character array. The text to be imported for training can be essays, poems, some kind of code, or even a data-set:). But, please make sure that this file contains at least 1MB of text. Otherwise, you might not like the results. I also considered word-level (check Text-word) training in which vocab consists of unique words in the given dataset. The rest process is similar to that in character-level training.

2. model.py file defines the LSTM model and its training hyper-parameters i.e., units in a layer, dropout, training epochs, batch size etc. With each epoch, it saves the updated weights while training with the lowest error so that it can just load and generate new text without training the model again and again.

3. generate.py file assumes that the model is trained and is ready to generate new texts. This file takes input as the number of new character that is to be generated.

Execution:

• Write this on terminal:

$ python main.py [your_training_file]

• It will train the model now.
• When done, check the generated text in [your_result_file]-out.txt.

Notes

• You can modify the RNN model as well as adjust its hyperparameters to improve the results.
• It's not necessary to train a *.txt file. You can also use *.csv or some kind other texts.
• You can check wiki.txt and train the model with it. It contains page contents concatenated into one single file.

Thanks.