added Generating fonts with adversarial networks example

[TIXFeniks]

added Generating fonts with adversarial networks example

[ebenolson]

Neat stuff! Have you played with doing a conditional version?

Do you know the license of the notmnist dataset? It would be best to rehost it in the Recipes S3 bucket if allowed, and maybe a dump of the trained model weights as well.

[TIXFeniks]

How can we rehost it in the s3 bucket? No, I haven't tried the conditional version yet. I will upload pre trained weights and update the notebook as soon as i get to computer.

[justheuristic]

As for conditional version, i did a tiny rework of @TIXFeniks 's notebook to make it 'conditional' and it does work:
https://gist.github.com/justheuristic/f627ae57e697b02dd8a7bb23d968e593

[justheuristic]

Could you please repeat how to upload to AWS?

[f0k]

I did a tiny rework of @TIXFeniks 's notebook to make it 'conditional' and it does work

That's cool. I think I'd prefer the conditional version if we are only keeping one of them. (Reworking the notebook so the conditioning is only added for an additional experiment in the end would be nice, but seems to be cumbersome.) We'd need some explanatory words, though. Before "Before we dive into training them", you could add some notes on (a) that you add conditional variables to make it easier to learn and to allow conditional sampling and (b) that G(z) and D(x) will thus change to G(z, y) and D(x, y). And the beginning of "Generative models 101" has a broken image.
Oh, and instead of wget http://yaroslavvb.com/upload/notMNIST/notMNIST_small.tar.gz -O notMNIST_small.tar.gz I'd suggest wget -N http://yaroslavvb.com/upload/notMNIST/notMNIST_small.tar.gz (the output filename is notMNIST_small.tar.gz by default, and -N will only download it if the remote files has changed).

Could you please repeat how to upload to AWS?

Only us maintainers can do so. We need to know which files to upload and then we can send you the link to include in your notebook.

Would you like to change this PR to include the conditional version? Or shall we add both? (I wouldn't mind having both Generating fonts with adversarial networks.ipynb and Generating fonts with conditional adversarial networks.ipynb, but I don't know what Eben thinks.)

[TIXFeniks]

Fixed what you asked. By the way, we've lost our GPU for a while. Could you train the networks yourself and upload pretrained weights?