recipeRNN

This is a small project to develop an RNN language model to generate recipes(for food). Included is the code needed to scrape training data get_pages.py and parse_html.py as well as for training the model itself recipe_rnn.py

Data

You can scrape html pages using the following(e.g. to get recipes 10000 to 20000):

python get_pages.py 10000 20000

Then, parse the raw html into a usable text file using the following hacky script:

python parse_html.py recipe_download_dir/

There is also a flag in parse_html.py to include instructions as well as ingredients, but these are a harder language modeling problem to solve and are not well handled currently.

Regardless, this generates a set of *_ingredients.txt files containing the recipe from each html page.

Embedding

You will need to train a word embedding before training the RNN. I used the GloVe repo and trained a GloVe embedding on the recipe corpus above, length of 100.

Before running GloVe, run the following to clean the text files and concatenate them for input to GloVe:

python recipe_rnn.py --concat recipe_concat.txt

Training

Once you've got the GloVe embedding ready, you can train the RNN like so:

python recipe_rnn.py

There are a number of default parameters and filepath arguments defined, and you may want to define them differently, use the --help flag to see all arguments.

To generate recipes using the trained RNN, add --load model.pt.weights to the command to load the pre-trained weights.

Acknowledgements

This project adapts code from Sebastian Seung's COS 485 Neural Networks: Theory and Applications course, as well as the Pytorch word-level language modeling RNN example.

Further, I'm grateful for the existence of online recipe websites like Allrecipes.com, without which this project would not be possible.