Update poem-generator

README.md

@@ -1,6 +1,6 @@
 # pmlg-poem-generator
 
-This is a poem generator trained with Gated Recurrent Units (GRU) on a corpus of poems including those in:
+This is a poem generator trained with Gated Recurrent Units (GRU) on a corpus of poems including:
 * [Poems Every Child Should Know](http://www.gutenberg.org/ebooks/16436), by Mary E. Burt
 * [Poems](http://www.gutenberg.org/ebooks/52546), by Thomas Hall Shastid
 * [Poems of Progress and New Thought Pastels](http://www.gutenberg.org/ebooks/3228), by Ella Wheeler Wilcox
@@ -12,20 +12,40 @@ Most of the code comes from [Tensorflow's text generation tutorial](https://gith
 
 Please refer to the notebook for an overview of the results.
 
+## How to generate poetry?
+
+Use the pretrained model called poem_generator.
+
+ - In the same folder, save the notebook called poem_generator.ipynb and the 2 npy files: model_poems.npy, and model_rhymes.npy 
+ - Run poem_generator.ipynb
+
+Disclaimer: Saving the model with npy files is probably less efficient than loading a Checkpoint, but I did not manage to do it
+
+
+## How to train your own model
+
+(optional) if you want to train a model on your own text files:
+
+ - Find some text, called the file 'poem_corpus.txt'
+ - Run corpus_cleaner.py. It should create two text files: 'corpus.txt', and 'rhymes.txt'.
+
+To train a model:
+
+ - Run 'rhyme_model.ipynb'. It creates 6 npy files that store the weigths of your rhyme model.
+ - Run 'poem_model.ipynb'
+ - Run 'poem_generator.ipynb'
+
+
 ## Improvement ideas
 
 ### Short term:
-* Clean the corpus
-* Find more data
-* Save a model
-* Try LSTM
-* Train a model based on sequences of words instead of characters
+* Update weights with another corpus
+* Try data from [GutenTag](http://www.cs.toronto.edu/~jbrooke/gutentag/)
+* Train a model based on sequences of words instead of characters (memory issues)
 
 ### Middle term goals:
-* Create a rhyme generator
-* Train a model backward: from the last word to the first
-
-The idea is to generate the rhymes first, and then write the lines from the last word to the first. This process makes sure that the generated poems rhyme.
+* Try Bi-directional models (failed to train a model conform to [Tensorflow's text generation tutorial](https://github.com/tensorflow/tensorflow/blob/r1.10/tensorflow/contrib/eager/python/examples/generative_examples/text_generation.ipynb).)
 
 ### Long term goals
+* Style transfert
 * Generate other kind of text (Haiku, magazines, articles, books...)

model_training/README.md

@@ -0,0 +1,13 @@
+# Train your own poem generator model
+
+There are notebooks of interest here:
+ * A poem generator called *poem_model*. It trains a model that generates text backwards (from the last letter to the first). The idea is to use it with a rhyme generator.
+ * A rhyme generator called *rhymes_model*. It trains a rhyme generator supposed to return lists of words that rhyme
+
+The first notebook relies on a corpus of poems named 'corpus.txt'
+
+The second notebook will take 'rhymes.txt' as a database of rhymes.
+
+Note that 'rhymes.txt' is nothing but the rhymes extracted from 'corpus.txt', with the help of corpus2rhymes.py. 
+
+Note also that corpus2rhymes.py performs some preliminary corpus cleaning as well. Note that it is not perfect.