Picnic Support

# Picnic Hackthon 2019

## Team The team consists of:

• Bas Niesink (https://devpost.com/basanaz)
• Christel Geurts (https://devpost.com/ChristelG)
• David Nauta (https://devpost.com/davidnauta)
• Kevin van den Broek (https://devpost.com/Kevintjeb)
• Mat Blom (https://devpost.com/matblom)
• Yoshua Nahar (https://devpost.com/yoshuaNahar)

We are a mix of Data Scientists and students, all active at Info Support.

Solution

We've created several models and 'stacked' the resulting submission files. The tsv-folder contains the tsv's of the individual and stacked models.

Two of the models were built using the excellent fast.ai-library. A third one uses Microsoft's Custom Vision service.

A fourth model, which isn't included, was built using Keras.

We've tried to create an OCR model as well, to take advantage of the text labels on the products, but the results were unsatisfactory.

Submissions

Three submissions were created:

• 1: The probabilities generated by all three models summed together
• 2: The probabilities of the local (non-cloud) models summed together
• 3: The product of the probabilities of all models

Reproducability

The results can be reproduces by executing the Jupyter Notebooks in order. The ZIP file containing the dataset needs to be placed in the same folder as the notebooks.

API keys are needed to reproduce the Custom Vision model's results. These keys can be acquired free of charge. Instructions are provided in the respetive notebook.