News
๐ค Transformers 4.6 adds VISION!
Transformers v4.6 is our first release dedicated to computer vision!
1๏ธโฃ CLIP from OpenAI, for Image-Text similarity or Zero-Shot Image classification
2๏ธโฃ ViT from GoogleAI
3๏ธโฃ DeiT from FacebookAI
Try SOTA image classification with ViT and DeiT on the Model Hub!
๐ค Datasets 1.6
๐คDatasets v1.6 brings you speed, features, and of course datasets:
- Now blazing fast: ~0.1ms per query for a 100 billion rows dataset ๐๐คฏ
- Even faster for small ones in memory by default ๐
- Easy datasets concatenation: row โ๏ธ, column โ๏ธ, from memory ๐ง or disk ๐ฝ
- 800+ datasets available ๐, now with CUAD, OpenSLR, GEM1.1 and more
๐ Expert Acceleration Program
This new program offers direct premium support from the Hugging Face team, to accelerate companies in their Transformers journey.
๐ฎ Which model to fine-tune, how?
๐ How do I reduce latency by 10X?
โ๏ธ How do I optimize my production setup?
๐ง How do I leverage Transformers in SageMaker?
๐ How do I mitigate bias in datasets and models?
Contact us to learn more!
๐ธ Inference API - now from $9/mo!
The Accelerated Inference API is now available through our $9/mo Supporter plan!
Itโs the easiest way to integrate and serve any of the 13,000+ Hugging Face models - or your own private models - using our accelerated and scalable infrastructure, via simple API calls.
๐ค AutoNLP - now with Speech Recognition!
Create and deploy fine-tuned state of the art models automagically with AutoNLP!
New this month:
๐ Summarization models
๐ฃ Speech Recognition (ASR) models
๐ Regression models
๐บ๐ณ New languages: Hindi, Japanese, Chinese and Dutch
Let us know which task or language you'd like us to add next!
๐ Welcome JAX!
GoogleAI's JAX/Flax library can now be used as Transformers' backbone ML library.
JAX/Flax makes distributed training on TPU effortless and highly efficient!
Over 3,000 pretrained model checkpoints have been converted to JAX and can be fine-tuned on Natural Language Understanding downstream tasks.
๐ Google Colab
๐ Runtime evaluation
๐ค Accelerate
Want to run your PyTorch training loop on multi-GPUs or TPUs without using an abstract class you can't control or tweak easily? Try out our new open source library, ๐ค Accelerate!
With just five lines of code to add, your script will run locally (for debugging) as well as on any distributed setup!
Community
๐ธ BigScience
Over 500 leading researchers from around the world are contributing to BigScience to create new understanding and shared scientific artifacts on how Large Language Models behave.
We are proud to start this collaboration, read all about it in this excellent Tech Review piece by Karen Hao!
๐๏ธWav2Vec2 Fine-Tuning Week
During the Wav2Vec2 Fine-Tuning Week 382 community members came together to democratize state-of-the-art speech recognition technology for over 70 languages - thank you!
Participants fine-tuned a pretrained wav2vec2-large-xlsr-53 checkpoint on a language of their choice. Overall, over 240 fine-tuned model checkpoints were uploaded to the Model Hub, with many of them marking a new state-of-the-art performance in a low-resource language.
๐จโ๐ฉโ๐งโ๐ฆ Member spotlight
This month we tip our hat to Vasudev Gupta who did an incredible job contributing Googleโs BigBird to Transformers. On behalf of the Hugging Face Community, thank you Vasu!
Vasu added both the auto-encoding model checkpoint, bigbird-roberta-base as well as the seq2seq model checkpoint, bigbird-pegasus.
Tutorials
๐ง Distributed Training on SageMaker
Learn how to use the new Hugging Face DLCs and Amazon SageMaker to train a distributed Seq2Seq-transformer model on the summarization task using the transformers and datasets libraries.
๐ Scaling BERT Inference on CPU
We partnered with Intel to uncover all the knobs to speed up inference on modern CPUs.
The result: a new library to dial-in and measure your inference setup, with an in-depth blog post to dig into the details.
Events & Talks
๐บ May 26th: PyTorch Community Voices
Transformers core maintainers Sylvain Gugger and Lysandre Debut boiled down the Hugging Face ecosystem and took live questions from the PyTorch community.