Philipp Schmid’s Post

Suraj Chawla of GPU.net leads the 2nd module in our 𝐅𝐑𝐄𝐄 "𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐀𝐈 𝐒𝐭𝐚𝐜𝐤 𝐟𝐨𝐫 𝐄𝐧𝐭𝐞𝐫𝐩𝐫𝐢𝐬𝐞 𝐋𝐞𝐚𝐝𝐞𝐫𝐬" 𝐜𝐨𝐮𝐫𝐬𝐞! 🚀 In this module, he covers the infrastructure layer, focusing on the hardware that powers AI. This module covers the importance of GPUs in processing large datasets and running complex AI models efficiently. 🤖 🧐 Want to know more about the different layers of the GenAI stack? Sign up for our course now to start learning! ✍🏼 https://lnkd.in/gpayDPdz

Excited to share my very first ML project! 🚗💻 Using YOLOv8, I built a car counter that accurately detects and tracks vehicles in videos. It was an incredible journey diving into the world of machine learning and computer vision. One major challenge I faced was converting the YOLO model, which uses PyTorch, to an OpenVINO model in order to integrate with my Intel GPU. Overcoming this hurdle taught me valuable lessons in optimization and deployment. Check out the demo video to see it in action! #MachineLearning #ComputerVision #YOLOv8 #CarCounter #MLProject #OpenVINO #ChallengeAccepted

Just finished the course “Image Processing” by MathWorks. https://lnkd.in/gGhq4pdN

🚀 Excited to share my work on Opencv ! 🎥👨💻 Just implemented real-time face detection, showcasing the power of computer vision! 🤖✨ It's incredible to see how just a few lines of code can turn a basic camera feed into an intelligent system that recognizes faces. This is just the beginning—looking forward to exploring more advanced applications! 💡🔍 It's just an basic use case of opencv, the scope of this library is Limitless. #ComputerVision #OpenCV #TechInnovation #FaceDetection

If you’re not taking advantage of quantization, you’re losing money. Quantized models have reduced memory footprint, can be ran with less resources and faster! And you can obtain all that using Hugging Face optimum. Using the optimum package you can quantize your transformer models with just a CLI command, and there are also different options designed for different device types (CPU or GPU). You can also use the same package to export your models to the classic ONNX format. And the best part: you can use pipelines to perform your inferences, getting those beautiful results that HuggingFace pipelines usually provide! If you enjoyed this, join me in exploring the world of programming, machine learning, and MLOps by hitting the following button. Let's level up together! 🔥🤖 #mlops #machinelearning

📄These days I've been learning and working on image processing and machine vision running on Raspberry Pi (CM4 and Pi 5). In my couple of weeks' journey, I noticed there are no MIPI camera examples out there for video streaming or Machine Learning applications. 🙇♂️ Why not?🤔 I guess that using USB cameras is simpler and cheaper... but with MIPI cameras as good as the new Raspberry Pi Camera Modules I got obsessed with using them instead of the USB ones. This is why I am sharing with you this example I developed showing a very simple way to stream live video using a MIPI camera and the Picamera2 library. 📁Link to Repo: https://lnkd.in/gzxDYdwP

#Nvidia's #BigVGAN v2, highlighting its enhanced features for audio processing. It details improvements such as a custom #CUDA kernel for faster inference on A100 #GPUs, an improved discriminator and loss function for better audio quality, and the inclusion of diverse audio types in training data. Additionally, it mentions the availability of permissive pre-trained model checkpoints that support higher sampling rates and upsampling ratios. #mldk #mldktech #mldkgenai https://mldk.tech

My final project for my machine learning module: Fast music videos from bad doodles. A LoRA was trained on some awful drawings by my housemates and I, which was then injected into stable diffusion. SDXl-turbo (announced days before making this) allowed me generate about a frame a second on an RTX 3060 GPU. For comparison, latent walks I’d done in the past took around a week to generate. Dataset by myself, Abby Bonds, Lizzie Allison, Maizie Rayne, and Teddy Clarke. Inspired by Nathan Raw, Mario Klingemann and EgadZoundsGadzooks. My code can be found here: https://lnkd.in/dKrV7NJF

🎉 Excited to share our latest work on extracting better features from diffusion models co-led by Stefan Baumann and Kolja Bauer. ✨ Diffusion models are amazing at learning world representations. Their features power many tasks: • Semantic correspondence • Depth estimation • Semantic segmentation 🤔 But have you ever wondered why we extract diffusion features from noisy images? Doesn’t that destroy valuable information? We show it does - and also requires finding correct hyperparameters for every downstream task. We thought, there had to be a better way. And there is. 🚀 With just 30 minutes of task-agnostic finetuning on a single GPU, we eliminate the need for noisy inputs. ✅ No noise ✅ No timestep tuning ✅ Better features, better performance across many tasks We make code and cleaned 🧹 weights available for SD 1.5 and SD 2.1. Have a look now! ⬇️

Nvidia released BigVGAN v2! 🎧 > Custom CUDA kernel for inference: w/ fused upsampling + activation kernel upto 3x faster inference on A100 > Improved discriminator and loss: using a multi-scale sub-band CQT discriminator and a multi-scale mel spectrogram loss > Larger training data: trained using datasets containing diverse audio types, including speech in multiple languages, environmental sounds, and instruments. > Permissive pre-trained model checkpoints supporting up to 44 kHz sampling rate and 512x upsampling ratio > Models & Demo on the Hub 🤗