call_for_contributions.md

Call for code example contributions

This is a constantly-updated list of code examples that we're currently interested in.

If you're not sure whether your idea would make a good code example, please ask us first!

---

Structured data examples featuring Keras Preprocessing Layers (KPL)

E.g. feature hashing, feature indexing with handling of missing values, mixing numerical, categorical, and text features, doing feature engineering with KPL, etc.

---

Transformer model for MIDI music generation

Reference TF/Keras implementation

---

StyleGAN2

• Paper
• Reference TF implementation

---

Text-to-speech

Example TF2/Keras implementation

---

Learning to rank

Reference Kaggle competition

---

DETR: End-to-End Object Detection with Transformers

• Reference implementation
• TF/Keras implementation

---

3D image segmentation

---

Question answering from structured knowledge base and freeform documents

---

Instance segmentation

• Tensorflow-YOLACT
• Additional references

---

EEG & MEG signal classification

---

Text summarization

---

Audio track separation

---

Audio style transfer

---

Timeseries imputation

---

Customer lifetime value prediction

---

Keras reproducibility recipes

---

Standalone Mixture-of-Experts (MoE) layer

MoE layers provide a flexible way to scale deep models to train on larger datasets. The aim of this example should be to show how replace the regular layers (such as Dense, Conv2D) with compatible MoE layers.

References:

• A relevant paper on MoE: https://arxiv.org/abs/1701.06538
• Switch Transformers on keras.io
• Keras implementation of Dense and Conv2D MoE layers

---

Guide to report the efficiency of a Keras model

It's often important to report the efficiency of a model. But what factors should be included when reporting the efficiency of a deep learning model? The Efficiency Misnomer paper discusses this thoroughly and provides guidelines for practitioners on how to properly report model efficiency.

The objectives of this guide will include the following:

• What factors to consider when reporting model efficiency?
• How to calculate certain metrics like FLOPS, number of examples a model can process per second (both in training and inference mode), etc?