fl-text-models

In this project we pretrain federated text models for next word prediction. See our white paper for details on related work, experiment designs, and results, or check out the video below for an overview of our research.

Data

The main dataset used for these experiments is hosted by Kaggle and made available through the tff.simulation.datasets module in the Tensorflow Federated API. Stack Overflow owns the data and has released the data under the CC BY-SA 3.0 license.

Environment

Experiments were conducted in a Python 3.7 conda environment with the packages in requirements.txt on both GPU and CPU VMs running Ubuntu 16.04.

Running Experiments

To conduct experiments with our code:

• Clone the repository and replicate our conda environment.
• Configure the params.json file to set a simulated client data sampling strategy, pretraining approach, and federated model architecture.
• Execute federated_nwp.py to train a federated text model on Stack Overflow for next word prediction according to the desired parameters. This script applies our methods described in final_research_report/README.md and is based on work from the research section of the Tensorflow Federated API.
• Model weights, train and validation statistics, plots, and client sample metadata are automatically stored in the experiment_runs directory. Run experiment_runs/training_progress.py to summarize model performance during or after training.
• See the notebooks directory for additional analysis, experiments, and examples of loading, testing, and comparing trained models.

Example Setup and Execution

git clone https://github.com/federated-learning-experiments/fl-text-models.git
conda create --verbose --yes --name tff python=3.7
conda activate tff
pip install -r requirements.txt
python federated_nwp.py

Due to the matplotlib dependency, you may need to apply the fix recommended here, if Python is not recognized as a framework.

References

This project draws mainly from the following research, but other sources are referenced throughout this repository, particularly code snippets. Special thanks to Keith Rush and Peter Kairouz from Google for their guidance throughout the course of this project.

• [1] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, Blaise Aguera y Arcas. “Communication-Efficient Learning of Deep Networks." Accessed December 6, 2019.
• [2] Keith Bonawitz, Hubert Eichner, Wolfgang Grieskamp, Dzmitry Huba, Alex Ingerman, Vladimir Ivanov, Chloe Kiddon, Jakub Konecny, Stefano Mazzocchi, H. Brendan McMahan, Timon Van Overveldt, David Petrou, Daniel Ramage, Jason Roselander. “Towards Federated Learning at Scale: System Design.” Accessed December 6, 2019.
• [3] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ramaswamy, Francoise Beaufays Sean Augenstein, Hubert Eichner, Chloe Kiddon, Daniel Ramage. “Federated Learning for Mobile Keyboard Prediction.” Accessed December 6, 2019.
• [4] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. "GloVe: Global Vectors for Word Representation." Accessed February 1, 2020.
• [5] Piotr Bojanowski, Edouard Grave, Armand Joulin, Tomas Mikolov. "Enriching Word Vectors with Subword Information." Accessed February 17, 2020.
• [6] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever. "Language Models are Unsupervised Multitask Learners." Accessed February 27, 2020.
• [7] Jiaqi Mu, Pramod Viswanath "All-but-the-Top: Simple and Effective Postprocessing for Word Representations." Accessed March 05, 2020.
• [8] Vikas Raunak, Vivek Gupta, Florian Metze. "Effective Dimensionality Reduction for Word Embeddings." Accessed March 05, 2020.
• [9] Vikas Raunak, Vaibhav Kumar, Vivek Gupta, Florian Metze. "On Dimensional Linguistic Properties of the Word Embedding Space." Accessed February 27, 2020.
• [10] Sashank J. Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konecný, Sanjiv Kumar, H. Brendan McMahan. "Adaptive Federated Optimization." Accessed March 6, 2020.
• [11] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Jamie Brew. "HuggingFace's Transformers: State-of-the-art Natural Language Processing." Accessed March 05, 2020.
• [12] S. Hochreiter, J. Schmidhuber. "Long Short-Term Memory." Accessed March 13, 2020.

Contact

• Arjun Singh*
• Joel Stremmel*