Recap of TensorFlow Dev Summit Extended Nairobi ‘19
Much like the excitement & anticipation surrounding the premiere of the Game of Thrones season finale, with the release of TensorFlow 2.0, much hype reverberated across the ML community all around the world. TensorFlow 2.0 is focused on ease of use, with APIs for beginners and experts to create machine learning models. Among the biggest features announced was that TF 2.0 would put Keras as the central high-level API used to build and train models. Keras provides several model-building APIs (Sequential, Functional, and Subclassing), so you can choose the right level of abstraction for your project.
On April 6th, 80+ TensorFlow enthusiasts from all around Nairobi converged at iHub to attend the TensorFlow Dev Summit Extended Nairobi, organized by GDG Nairobi and partnering with communities such as Nairobi AI. The event encapsulated all the major announcements made during the parent event held in Sunnyvale, CA on March 6 and 7, 2019. With presentations from an amazing speaker line-up, attendees were treated to top-notch demos and codelabs, further showcasing how TensorFlow helping to power the ML revolution. Here’s a brief of the sessions carried out.
Keynote
Speakers: Ngesa Marvin, Irene Onyango
GDG Nairobi organizers Irene Onyango and Ngesa Marvin, kicked off the event by engaging the crowd on pre-keynote activities. Ngesa trumpeted how TensorFlow has grown to be the most widely used ML platform in the AI community. He emphasized how the onset of TensorFlow 2.0 was an exciting milestone, bringing even greater features and more abstractions to ML developers.
Codelab: Fine Tuning Language Models with TensorFlow
Speaker: Victor Akinwande, IBM Research Africa
TensorFlow provides numerous capabilities to implementing and fine-tuning language models. Victor Akinwande from IBM Research Africa engaged the audience in a codelab session on how to perform sentiment analysis using BERT (Bidirectional Encoder Representations from Transformers). Given NLP is fast becoming the main user interface for most AI systems, researchers from the Google AI Language team open sourced BERT, a new library for pre-training language-representations and achieving state-of-the-art performance.
Victor dived deep into the intricacies of BERT, including its bidirectional contextual representations. Sketching out the model’s architecture at the front, he explained the two stages if using BERT: pre-training and fine-tuning.
To learn more on BERT, here’s a comprehensive blog article: BERT in Keras with TensorFlow Hub
TensorFlow Hub
Speaker: Chris Barsolai, Nairobi AI
The main focus of TensorFlow is to make machines intelligent and improve people’s lives. Their goal is to make ML more accessible by everyone. A major push forward in this vision is abstracting the numerous technicalities surrounding ML systems in production.
Leveraging from the concept of how shared repositories of code through platforms like GitHub make us more productive by letting programmers quickly reuse code, TensorFlow aims to create this same foundational shift in ML. Barsolai illustrated how TF-Hub makes it easy to create, discover, publish, share and resuse pieces of ML modules in TF. He facilitated an introduction to the concept of pretrained building blocks called modules, and transfer learning for image retraining and text classification.
Barsolai pointed out how TensorFlow Hub was a tool that could highly benefit African researchers and ML engineers, by allowing them to create highly accurate models using less data and less domain expertise.
Demo: TensorFlow for billboard management
Speaker: Bildad Namawa, JUMO
Central to the event was showcasing projects built by the community and powered by TensorFlow. For this session, Bildad Namawa from JUMO World provided an in-depth walk-through into a project he was working on. The project featured billboard management through object detection from images using TensorFlow.
Bildad illustrated how he’d gone about collecting data using Google Street View, then after obtaining a large enough dataset, trained the model over 56k steps until the model achieved a satisfactory accuracy score. With additional suggestions from the crowd, he listed the various use cases in which the billboard identification from imagery could be used e.g. revenue collection by county governments. It was easy to see how much the audience was impressed by the project, owing to the large number of questions asked.
TensorFlow implementation of a GAN Network
Speaker: Allan Chepkoy, iCodeAI
Generative adversarial networks (GANs) are one of the greatest advances in AI and ML. GANs, first introduced in a paper by Ian Goodfellow in 2014, are designed to mimic any distribution of data. They can be taught to generate fake data in any domain.
Allan Chepkoy took the audience through implementing GAN networks using TensorFlow. He demoed how to use the MNIST data set to play around with a simple GAN made using Tensorflow’s layers API.
Codelab: Applying advanced ML on images
Speaker: Tonny Muthiri
Presently, there exists several CNNs trained models using various CNN architectures such as LeNet, AlexNet, VGG, GoogLeNet Inception and ResNet. Tonny took the audience through ways of applying advanced ML on images using TensorFlow.
CIFAR-10 classification is a common benchmark problem in machine learning. Tonny demonstrated how to train a model on your own dataset, including data augmentation technique to solve dataset limitation by transform.
The TensorFlow Party Continues
Undoubtedly, the onset of TensorFlow 2.0 ushers in more capabilities for ML researchers. We can’t wait to see all the amazing solutions ML enthusiasts build with the aim of solving challenging problems with AI, more so in Africa.
It doesn’t end here. We recently announced TensorFlow Nairobi, in a spirited effort to join the global TensorFlow community. We will be hosting several meetups over the coming months led by experts, each targeted at a core TensorFlow feature and functionality. To join our community, sign up on our Meetup page (https://meetup.com/TensorFlow-Nairobi), & follow us on Twitter for instant updates to everything TensorFlow.
If you still haven’t joined the community, what are you waiting for? It won’t be the same without you. We’ll be seeing you at TF Nairobi!
Lots of thanks to GDG Nairobi through the organizers Ngesa Marvin and Irene Onyango, who put together such an awe-inspiring event. We hope to see more of AI-oriented events from community.
Follow the Nairobi AI publication and join our community on Meetup to see more updates and stories on the state of AI and ML in Nairobi and Africa.
