Piotr Skalski’s Post

Fine-tune Florence-2 on custom object detection dataset 🔥 🔥 🔥 (super technical blog post; video tutorial coming this week) TL;DR: learn how to fine-tune powerful Florence-2 VLM in Google Colab Florence-2 is a lightweight vision-language model open-sourced by Microsoft (Bin Xiao, Haiping Wu, Weijian Xu) under the MIT license. It excels in tasks like captioning, object detection, grounding, and segmentation. Like other pre-trained foundational models, Florence-2 may lack domain-specific knowledge and may need fine-tuning. This tutorial covers: - Dataset format required for training - Load pre-trained Florence-2 model - Run object detection, captioning, and grounding tasks - Configure LoRA for optimized training (allowing you to train larger models on a single GPU) - Train and benchmark fine-tuned model Huge thanks to Andrés Marafioti and Peter Robicheaux for their help in creating the code used in this tutorial! 🙏🏼 ⮑ 🔗 blog post: https://lnkd.in/dp5bUt4m #computervision #objectdetection #transformers #multimodalai #opensource

Introducing OpenLS-DGF: a framework enhancing ML tasks in logic synthesis with adaptive dataset generation. 🚀 Explore diverse downstream tasks in circuit design and optimization with OpenLS-D-v1. #DataScience #MachineLearning #LogicSynthesis 🔍

🔎 Unlocking the Power of Image Analysis: Introducing Embeddings🔍 Have you ever wondered how computers make sense of complex images in scientific research? Enter embeddings - a game-changing tool in data science that's revolutionizing image analysis. Embeddings transform high-dimensional data, like images, into compact vectors of numbers. This allows computers to efficiently process and analyze visual information, moving beyond traditional pixel-by-pixel approaches. At Spring Science, we've harnessed the potential of embeddings in our Spring Engine platform. By converting images into these numerical representations, we're enabling faster, more accurate analysis of cellular phenotypes in high-content screening. Why does this matter? Embeddings are paving the way for more efficient drug discovery and deeper understanding of biological processes. They're helping scientists uncover patterns and insights that were previously hidden in the complexity of image data. Are you curious about how embeddings could transform your research? Let's dive deeper into this fascinating world of data science and image analysis. #ImageAnalysis #DataScience #ScientificResearch https://buff.ly/4dLZ6jQ

📣 Big Announcement! 📣 We are thrilled to introduce LLMWhisperer, Unstract's cutting-edge text extraction service! 🌟 ➡ https://lnkd.in/gusJH5Ui Unlocking the power of advanced text extraction, LLMWhisperer effortlessly deciphers complex documents with precision, tailored for optimal understanding by LLMs. 💡 Ready to witness its capabilities firsthand? Dive into the action with our exclusive 100 pages/day, free forever plan. 📚✨ ➡ https://lnkd.in/gVieTjCv Explore the magic yourself in the LLMWhisperer Playground, where testing with your own documents is a breeze - and no sign-up required! ➡ https://lnkd.in/gAVTUT8Y ✨ Advanced capabilities of LLMWhisperer ❇ Layout Preserving Mode: this helps preserve the layout of input documents, making it easy for LLMs to extract data from complex documents with multiple columns or tables. ❇ Checkbox and Radiobutton detection: It can convert these into text in a manner LLMs are able to interpret. Deal easily with those pesky forms that feature such elements. ❇ Auto Mode Switching: LLMWhisperer can switch automatically to OCR mode should text extraction fail on scanned documents, for example. ❇ Auto-Compaction: outputs as few tokens as possible, to save you LLM costs. #LLMs #ParseData #DataExtraction #DocumentParser

Unstract introduces LLMWhisperer: advanced text extraction for LLMs, with features including Layout Preserving Mode, Checkbox detection, Auto Mode Switching, and Auto-Compaction.

Advancing time series analysis with multi-granularity guided diffusion model; An algorithm-system co-design for fast, scalable MoE inference; What makes a search metric successful in large-scale settings; learning to solve PDEs without simulated data. https://msft.it/6045lyVfD

From Microsoft Research: GraphRAG (Graphs + Retrieval Augmented Generation) is a technique for richly understanding text datasets by combining text extraction, network analysis, and LLM prompting and summarization into a single end-to-end system -

🌟 Exciting news! Check out this insightful blog post on "Graph Vertex Embeddings: Distance Regularization and Community Detection" which delves into the fascinating world of graph embeddings and their impact on complex network structures. The post explores the use of efficient methods leveraging the metric structure in the embedding space as a proxy for the topological structure of the data. Dive into the details and explore competitive results with reduced computational complexity. Read the full post here: https://bit.ly/4aAGJxa #GraphEmbeddings #DataScience #CommunityDetection

🚀 𝐌𝐲 𝐉𝐨𝐮𝐫𝐧𝐞𝐲 𝐭𝐨 𝐑𝐞𝐩𝐥𝐢𝐜𝐚𝐭𝐞 𝐂𝐨𝐧𝐯𝐍𝐞𝐗𝐭 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 🚀 Last week, I achieved a significant milestone by successfully replicating the ResNet architecture. This week, I decided to dive into replicating the ConvNeXt architecture. 𝐋𝐢𝐧𝐤𝐬: 🔗 My replication notebook: https://lnkd.in/gwAkdCAk 📄 ConvNeXt research paper: https://lnkd.in/ggpEsVBK 𝐖𝐡𝐚𝐭’𝐬 𝐂𝐨𝐧𝐯𝐍𝐞𝐗𝐭? ConvNeXt is a cutting-edge computer vision architecture that bridges the gap between traditional ConvNets and modern vision transformers. Here’s what makes it special: 𝐇𝐲𝐛𝐫𝐢𝐝 𝐃𝐞𝐬𝐢𝐠𝐧: ConvNeXt combines the best of both worlds. It retains the straightforward, fully-convolutional structure of ResNets while incorporating key innovations inspired by transformers. 𝐄𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲: ConvNeXt maintains the simplicity and efficiency of ConvNets, making it suitable for various computer vision tasks. 𝐃𝐞𝐩𝐭𝐡𝐰𝐢𝐬𝐞 𝐂𝐨𝐧𝐯𝐨𝐥𝐮𝐭𝐢𝐨𝐧: ConvNeXt uses depthwise convolution, allowing it to capture global context efficiently. 🌟 𝐄𝐱𝐜𝐢𝐭𝐢𝐧𝐠 𝐓𝐢𝐦𝐞𝐬 𝐀𝐡𝐞𝐚𝐝 🌟 As I continue my journey, I’m thrilled to explore the ever-evolving field of machine learning. Stay tuned for more updates! 🤖📸 #AI #ComputerVision #DeepLearning #ConvNeXt #ResNet

Looking for a robust and automated method for detecting water from optical satellite imagery using Google #EarthEngine? We recommend an automated approach using unsupervised learning that combines multiple water indices and automatically label water pixels. This approach is based on the 'waterdetect' algorithm published at https://lnkd.in/geNmJ9Jh Check out the video explanation → https://lnkd.in/g-cQx5ah Access the Code → https://lnkd.in/gNmzydbA