1 of 16

Recipe Generator

Lucy Li & Hayley Bergman

LA: Isaac Kontomah

Industry Mentor: Feitong Yang

2 of 16

Intro

Running out of ideas in the kitchen? Need to spice things up when cooking? Look no further!

​

​

​

Goal: Develop a model that generates new unique recipes given an existing recipe dataset

3 of 16

What is an RNN?

  • Recurrent Neural Networks
  • Deep learning model for remembering information that occurs over time

​

prediction

new information

4 of 16

What is an LSTM?

  • An RNN with memory

prediction + memories

memory

forgetting

new information

prediction

5 of 16

What is a GRU?

  • Gated Recurrent Unit
  • Get to decide whether to pass previous output to the next cell or not

6 of 16

Motivation - Use of RNNs

Why use a RNN/LSTM?

  • Established approach to generation problems
  • Ability to generate text that will look like the original training data

​

7 of 16

Data

  • Epicurious recipe dataset
  • https://www.kaggle.com/hugodarwood/epirecipes.

8 of 16

Candidates

  • Character RNN vs Word RNN
    • Training Time
    • Coherency of sample results

​

9 of 16

Training Methodology

  1. Pre-process recipe dataset
  2. Select data to train model on
  3. Generate sample results

10 of 16

Training the Model

Using the Torch Neural Network Package

Input: data, RNN size, N, Dropout

​

Input LSTM..n times..LSTM output

11 of 16

Results

LSTM

12 of 16

Results

GRU

13 of 16

Results

RNN

14 of 16

Results

Word RNN

15 of 16

Lessons Learned

  • Deep learning process
    • Data gathering
    • Data Processing
    • Training Model
    • Adjusting Parameters
    • Interpreting Results
  • Finding and using resources

16 of 16

Future Work

  • Nutrition facts, cuisine type, etc.
  • GANs?
  • Recipe accuracy