Can Stable Diffusion be run on CPUs? - January 17, 2023

Intel shared a Hugging Face Spaces demo comparing the latency of Stable Diffusion on their newest XEON CPU (Sapphire Rapids) with the previous generation (Ice Lake), achieving 6x performance speedup and potentially enabling Stable Diffusion image generation and other Transformers use cases on CPUs. You can try the demo yourself on Hugging Face.

News & Announcements πŸ“£

CarperAI released CHEESE, a new open-source library to create datasets for Reinforcement Learning with Human Feedback easily.

Microsoft Research announced VALL-E, a language model for text-to-speech generation, with unimaginable results. VALL-E can synthesize high-quality personalized speech with only a 3-second recording of an unseen speaker as an acoustic prompt. Check out the examples here.

AllenAI published SODA, the first publicly available, million-scale, high-quality dialogue dataset covering a wide range of social interactions.

Hugging Face added the Summarize from Feedback an OpenAI RLHF dataset, to their Hub.

Niels added GIT by Microsoft to Hugging Face Transformers. GIT is an image-to-text transformers achieving state-of-the-art performance in image captioning and visual question answering.

Hugging Face silently made HFFS public. HFFS implements a fsspec for the Hugging Gace Hub to provide a convenient Python filesystem interface toΒ πŸ€—Β Hub. ****

Tutorials & Demos πŸ“

I published a blog post on how to get started with Transformers and TPU using PyTorch on Google Cloud. The blog posts cover an end-to-end example of how to fine-tune BERT on text classification.

The One Shot Talking Face demo allows you to convert an image and audio to a video of the person speaking.

DairAI created a new repository on ML Papers Explained including over 20 Transformer papers.

Monica Colangelo wrote a blog post on the 4 ultimate reasons to prefer AWS CDK over Terraform.

Sebastian Raschka shared some quality excerpts from his Deep Learning Fundamentals course about Embedding Layers.

Reads & Papers πŸ“š

Document UnderstanDing of Everything (DUDE), a new Document Understanding benchmark, was published to foster research and evaluate document understanding in a real-world setting.

Leo Boytsov created an awesome thread about inPArs-Light, a cost-effective unsupervised training method of efficient rankers for information retrieval and search.

If you are using Google Cloud and Terraform, you should look at building your perfect Google Cloud infrastructure using Terraform and the gcloud CLI.

Peter Izsak wrote an awesome blog article on positional encodings in language models and if they are Required.

Tim Dettmers updated his blog post on how which GPU(s) to Get for Deep Learning: My Experience and Advice for Using GPUs in Deep Learning

Shiye Lei and Dacheng Tao published a Comprehensive Survey of Dataset Distillation. Dataset

I hope you enjoyed this newsletter. πŸ€— If you have any questions or are interested in collaborating, feel free to contact me on Twitter or LinkedIn.

See you next week πŸ‘‹πŸ»πŸ‘‹πŸ»