November 2022 Summaries

Fine tuning is a common practice in deep learning that involves training a model on a specific dataset to improve its performance. In this example, the Stable Diffusion model was fine tuned on a Naruto anime character dataset to create a text-to-image model that generates custom Naruto-inspired images based on any text prompt. The trained model can produce high-quality images with characteristic elements of Naruto costumes. Prompt engineering plays a crucial role in producing compelling and consistent results. Using specific prompts, such as "person_name ninja portrait" or "person_name in the style of Naruto", can significantly improve the quality of the generated images. The model was trained on a large dataset using 2 GPUs for around 12 hours at a cost of about $20. The fine-tuned model is now available for use by others, and a live demo is provided to generate your own Naruto-like images.

The text discusses the launch of Lambda Reserved Cloud Clusters, which are designed for machine learning teams that require large compute power to train transformer models and other large datasets. The clusters are made up of 8x NVIDIA A100 bare metal GPUs connected with 1600 Gbps RDMA networking, offering faster node-to-node bandwidth than AWS's equivalent EC2 offering. This allows for multi-node distributed training, which is essential for creating large transformer or diffusion models. Voltron Data, a data company developing standard solutions for the Apache Arrow ecosystem, has partnered with Lambda to use these clusters for their needs. The clusters offer several benefits, including 25-57% savings on compute costs and 90% savings on egress fees compared to AWS. They also have flexible contract lengths starting at 6 months, dedicated GPUs, and fast networking speeds of 1600 Gbps. Additionally, Lambda provides pre-configured cloud products designed for deep learning, with support options available for premium customers.