A small proof-of-concept joke project: given some user-generated text, use a Markov chain, or a Recurrent Neural Net to intermix it with pre-existing data, and generate some new text.

This repository contains a small Meteor webapp (UI), and two backends (markov-backend) and (neural-backend). We use AWS Lambda for computation (it is practically free, but suboptimal, since it doesn't have GPU support). The backend directories contain code to generate the text, model (if applicable), and then to tar up and upload the code + model into AWS Lambda.

For control, we use the settings file -- there is a sample sample-settings.json in the root directory.

To make AWS Lambda calls, fill in the key and secret keys with a key & secret for an AWS IAM user authorized to run/change AWS Lambda code on your AWS account.

The neural-backend directory contains some python/chainer code that runs a neural net to, first, generate a model based on an author, use that model to generate samples, and remix the results with the user's input. The net logic itself is based off: [[https://github.com/yusuketomoto/chainer-char-rnn]] The zipped base_lambda.zip file contains the project's binary dependencies, as compiled and zipped from an AWS machine.

To add a new source:

• Choose an author-token (ex: "tolkien") and lamnda-name (ex: "tolkien-net"). The token is what we will use as a base filename for plaintext, model, etc. The lambda-name should be the name of the AWS Lambda function that you should define using the AWS API. Add these to the settings.json file, in the net-mixins stanza, as approporiate.

• Add a plaintext file of the author's work to data/<token>.txt.

• In the neural-backend directory, run generate_model.py <token> to generate a new model. The script will also use that model to generate a small sample, just to give some visibility in what that model produces.

• Once you are satisfied with the model, in the neural-backend directory, run generate_lambda.py <path to settings.json> <token> to publish the AWS Lambda.

To run the webapp with the neural net backend, set the mode in settings.json to net.

The markov-backend directory contains some python to run a simple markov chainer autocompleter. By default, it runs/updates in an AWS Lambda function called "markov-chainer", but the name could be changed in webapp/server/markov.js and markov-backend/generate_lambda.py.

To add a new source:

• Add a file called <token>.txt to the data directory. Add it to the markov-mixins in the settings file, as approporiate.

• In the markov-backend directory, run generate_lambda.py <path to settings.json>

To run the webapp with the markov backend, set the mode in settings.json to markov.

The webapp is a small UI that calls up the AWS Lambda functions on the server. You can run it with:

meteor --settings <settings.json>

You can deploy it for free with:

meteor deploy --settings <settings.json> <domain-name>.meteor.com