The Dreamup is an idea for a codeless alternative to the web. This repo is an initial exploration into using a tensor2tensor nlp model to produce HTML from a Dreamup application.
Status: This is in progress. Nothing is really working yet. There isn't yet a sufficient language->html dataset to feed our model. As it stands, the model prints out nonsense when given the nlp input. I'm looking at options of generating a valid dataset.
The ml-model directory contains the language->html tensor2tensor problem. It is setup to be run on Floydhub.
The data directory contains:
- An example seq2seq dataset. The
t2t-modeldir contains a tensor2tensor problem which generates similar seq2seq data as well. - hand written examples of nlp->html which are meant to server as example dreamup applications. These examples are not yet consumed by any model.
The browser dir is a super simple Dreamup browser. The apps dir are sample applications that you can point the dreamup browser at.
This repo is meant to be run on floydhub.com. The 'launch_floyd_run.sh' file will execute the full datagen-train-evaluate process. You will first need to setup your own floydhub project and purchase GPU credits. The environments and parameters can already be found in the code here.
Some possible areas to work on:
- producing a language->html dataset
- explore other models than seq2seq, like seq2tree
- implementing the server/client of the browser
- exploring different implementations all together (not nlp -> html)