Q1 ‘23 highlights and achievements

Posted by Nari Yoon, Bitnoori Keum, Hee Jung, DevRel Group Supervisor / Soonson Kwon, DevRel Program Supervisor

Let’s discover highlights and accomplishments of huge Google Machine Studying communities over the primary quarter of 2023. We’re enthusiastic and grateful about all of the actions by the worldwide community of ML communities. Listed here are the highlights!

ML Campaigns

ML Group Dash

ML Group Dash is a marketing campaign, a collaborative try bridging ML GDEs with Googlers to supply related content material for the broader ML neighborhood. All through Feb and Mar, MediaPipe/TF Advice Dash was carried out and 5 tasks have been accomplished.

ML Olympiad 2023

ML Olympiad is an related Kaggle Group Competitions hosted by ML GDE, TFUG, Third-party ML communities, supported by Google Builders. The second, ML Olympiad 2023 has wrapped up efficiently with 17 competitions and 300+ contributors addressing vital problems with our time – variety, environments, and many others. Competitors highlights embody Breast Most cancers Prognosis, Water High quality Prediction, Detect ChatGpt solutions, Guarantee wholesome lives, and many others. Thanks all for collaborating in ML Olympiad 2023!

Additionally, “ML Paper Studying Golf equipment” (GalsenAI and TFUG Dhaka), “ML Math Golf equipment” (TFUG Hajipur and TFUG Dhaka) and “ML Examine Jams” (TFUG Bauchi) have been hosted by ML communities around the globe.

Group Highlights

Varied methods of serving Steady Diffusion by ML GDE Chansung Park (Korea) and ML GDE Sayak Paul (India) shares the right way to deploy Steady Diffusion with TF Serving, Hugging Face Endpoint, and FastAPI. Their different challenge Advantageous-tuning Steady Diffusion utilizing Keras offers the right way to fine-tune the picture encoder of Steady Diffusion on a customized dataset consisting of image-caption pairs.

Serving TensorFlow fashions with TFServing by ML GDE Dimitre Oliveira (Brazil) is a tutorial explaining the right way to create a easy MobileNet utilizing the Keras API and the right way to serve it with TF Serving.

Advantageous-tuning the multilingual T5 mannequin from Huggingface with Keras by ML GDE Radostin Cholakov (Bulgaria) exhibits a minimalistic method for coaching textual content technology architectures from Hugging Face with TensorFlow and Keras because the backend.

Lighting up Photos within the Deep Studying Period by ML GDE Soumik Rakshit (India), ML GDE Saurav Maheshkar (UK), ML GDE Aritra Roy Gosthipaty (India), and Samarendra Sprint explores deep studying strategies for low-light picture enhancement. The article additionally talks a couple of library, Restorers, offering TensorFlow and Keras implementations of SoTA picture and video restoration fashions for duties akin to low-light enhancement, denoising, deblurring, super-resolution, and many others.

How one can Use Cosine Decay Studying Price Scheduler in Keras? by ML GDE Ayush Thakur (India) introduces the right way to appropriately use the cosine-decay studying fee scheduler utilizing Keras API.

Implementation of DreamBooth utilizing KerasCV and TensorFlow (Keras.io tutorial) by ML GDE Sayak Paul (India) and ML GDE Chansung Park (Korea) demonstrates DreamBooth approach to fine-tune Steady Diffusion in KerasCV and TensorFlow. Coaching code, inference notebooks, a Keras.io tutorial, and extra are within the repository. Sayak additionally shared his story, [ML Story] DreamBoothing Your Means into Greatness on the GDE weblog.

Focal Modulation: A substitute for Self-Consideration by ML GDE Aritra Roy Gosthipaty (India) shares a Keras implementation of the paper. Usha Rengaraju (India) shared Keras Implementation of NeurIPS 2021 paper, Augmented Shortcuts for Imaginative and prescient Transformers.

Photos classification with TensorFlow & Keras (video) by TFUG Abidjan defined the right way to outline an ML mannequin that may classify photos in keeping with the class utilizing a CNN.

Palms-on Workshop on KerasNLP by GDG NYC, GDG Hoboken, and Stevens Institute of Expertise shared the right way to use pre-trained Transformers (together with BERT) to categorise textual content, fine-tune it on customized information, and construct a Transformer from scratch.

Steady diffusion instance in an android utility — Half 1 & Half 2 by ML GDE George Soloupis (Greece) demonstrates the right way to deploy a Steady Diffusion pipeline inside an Android app.

AI for Artwork and Design by ML GDE Margaret Maynard-Reid (United States) delivered a short overview of how AI can be utilized to help and encourage artists & designers of their inventive house. She additionally shared a couple of use circumstances of on-device ML for creating creative Android apps.

Finish-to-Finish Pipeline for Segmentation with TFX, Google Cloud, and Hugging Face by ML GDE Sayak Paul (India) and ML GDE Chansung Park (Korea) mentioned the essential particulars of constructing an end-to-end ML pipeline for Semantic Segmentation duties with TFX and numerous Google Cloud companies akin to Dataflow, Vertex Pipelines, Vertex Coaching, and Vertex Endpoint. The pipeline makes use of a customized TFX element that’s built-in with Hugging Face Hub – HFPusher.

Prolong your TFX pipeline with TFX-Addons by ML GDE Hannes Hapke (United States) explains how you need to use the TFX-Addons parts or examples.

Textual Inversion Pipeline for Steady Diffusion by ML GDE Chansung Park (Korea) demonstrates the right way to handle a number of fashions and their prototype purposes of fine-tuned Steady Diffusion on new ideas by Textual Inversion.

Operating a Steady Diffusion Cluster on GCP with tensorflow-serving (Half 1 | Half 2) by ML GDE Thushan Ganegedara (Australia) explains the right way to arrange a GKE cluster, the right way to use Terraform to arrange and handle infrastructure on GCP, and the right way to deploy a mannequin on GKE utilizing TF Serving.

Scalability of ML Functions by TFUG Bangalore targeted on the challenges and options associated to constructing and deploying ML purposes at scale. Googler Joinal Ahmed gave a chat entitled Scaling Giant Language Mannequin coaching and deployments.

Discovering and Constructing Functions with Steady Diffusion by TFUG São Paulo was for people who find themselves excited by Steady Diffusion. They shared how Steady Diffusion works and confirmed an entire model created utilizing Google Colab and Vertex AI in manufacturing.

In Equity & Ethics In AI: From Journalism, Medication and Translation, ML GDE Samuel Marks (United States) mentioned accountable AI.

In The brand new age of AI: A Convo with Google Mind, ML GDE Vikram Tiwari (United States) mentioned accountable AI, open-source vs. closed-source, and the way forward for LLMs.

Accountable IA Toolkit (video) by ML GDE Lesly Zerna (Bolivia) and Google DSC UNI was a meetup to debate moral and sustainable approaches to AI improvement. Lesly shared in regards to the “ethic” aspect of constructing AI merchandise in addition to studying about “Accountable AI from Google”, PAIR guidebook, and different experiences to construct AI.

Ladies in AI/ML at Google NYC by GDG NYC mentioned scorching matters, together with LLMs and generative AI. Googler Priya Chakraborty gave a chat entitled Privateness Protections for ML Fashions.

Environment friendly Process-Oriented Dialogue Techniques with Response Choice as an Auxiliary Process by ML GDE Radostin Cholakov (Bulgaria) showcases how, in a task-oriented setting, the T5-small language mannequin can carry out on par with present programs counting on T5-base and even larger fashions.

Studying JAX in 2023: Half 1 / Half 2 / Livestream video by ML GDE Aritra Roy Gosthipaty (India) and ML GDE Ritwik Raha (India) coated the ability instruments of JAX, particularly grad, jit, vmap, pmap, and likewise mentioned the nitty-gritty of randomness in JAX.

In Deep Studying Mentoring MILA Quebec, ML GDE David Cardozo (Canada) did mentoring for M.Sc and Ph.D. college students who’ve pursuits in JAX and MLOps. JAX Streams: Parallelism with Flax | EP4 by David and ML GDE Cristian Garcia (Columbia) explored Flax’s new APIs to help parallelism.

March Machine Studying Meetup hosted by TFUG Kolkata. Two classes have been delivered: 1) You do not know TensorFlow by ML GDE Sayak Paul (India) offered some under-appreciated and under-used options of TensorFlow. 2) A Information to ML Workflows with JAX by ML GDE Aritra Roy Gosthipaty (India), ML GDE Soumik Rakshit (India), and ML GDE Ritwik Raha (India) delivered on how one may consider utilizing JAX useful transformations for his or her ML workflows.

A paper overview of PaLM-E: An Embodied Multimodal Language Mannequin by ML GDE Grigory Sapunov (UK) defined the small print of the mannequin. He additionally shared his slide deck about NLP in 2022.

An annotated paper of On the significance of noise scheduling in Diffusion Fashions by ML GDE Aakash Nain (India) outlined the results of noise schedule on the efficiency of diffusion fashions and methods to get a greater schedule for optimum efficiency.

Three tasks have been awarded as TF Group Highlight winners: 1) Semantic Segmentation mannequin inside ML pipeline by ML GDE Chansung Park (Korea), ML GDE Sayak Paul (India), and ML GDE Merve Noyan (France), 2) GatedTabTransformer in TensorFlow + TPU / in Flax by Usha Rengaraju, and three) Actual-time Object Detection within the browser with YOLOv7 and TF.JS by ML GDE Hugo Zanini (Brazil).

Constructing rating fashions powered by multi-task studying with Merlin and TensorFlow by ML GDE Gabriel Moreira (Brazil) describes the right way to construct TensorFlow fashions with Merlin for recommender programs utilizing multi-task studying.

Constructing ML Powered Internet Functions utilizing TensorFlow Hub & Gradio (slide) by ML GDE Bhavesh Bhatt (India) demonstrated the right way to use TF Hub & Gradio to create a completely useful ML-powered internet utility. The presentation was held as a part of an occasion known as AI Evolution with TensorFlow, masking the basics of ML & TF, hosted by TFUG Nashik.

create-tf-app (repository) by ML GDE Radostin Cholakov (Bulgaria) exhibits the right way to arrange and keep an ML challenge in Tensorflow with a single script.

Creating scalable ML options to help huge techs evolution (slide) by ML GDE Mikaeri Ohana (Brazil) shared how Google will help huge techs to generate affect by way of ML with scalable options.

Search of Brazilian Legal guidelines utilizing Dialogflow CX and Matching Engine by ML GDE Rubens Zimbres (Brazil) exhibits the right way to construct a chatbot with Dialogflow CX and question a database of Brazilian legal guidelines by calling an endpoint in Cloud Run.

Steady Diffusion Finetuning by ML GDE Pedro Gengo (Brazil) and ML GDE Piero Esposito (Brazil) is a fine-tuned Steady Diffusion 1.5 with extra aesthetic photos. They used Vertex AI with a number of GPUs to fine-tune it. It reached Hugging Face prime 3 and greater than 150K folks downloaded and examined it.