SkalskiP Profile picture
Open-source Lead @roboflow. VLMs. GPU poor. Dog person. Coffee addict. Dyslexic. | GH: https://t.co/dEmzMDGq5H | HF: https://t.co/4Lx1Yw34W7

Jul 1, 2024, 10 tweets

Florence-2 fine-tuning YouTube tutorial is finally out! (sorry it took me so long)

- running the pre-trained model with different vision tasks
- configuring LoRA
- training and benchmarking
- Florence-2 vs. top vision model

link:

↓ key takeaways

deep dive into the dataset format you'll need for Florence-2 object detection fine-tuning

starting this week all datasets on @roboflow Universe can be downloaded in a format compatible with Florence-2

using PEFT library to configure LoRA for Florence-2 fine-tuning

defining training loop and fine-tuning on custom dataset

interesting to see that fine-tuned Florence-2 can accidentally produce misspelled class names

in the end I got mAP = 0.75 with Florence-2 vs. mAP = 0.91 with YOLOv8 (on the same dataset)

Florence-2 vs. other top computer vision models right now

here's my Florence-2 overview blog post if you want to learn more about the model

link: blog.roboflow.com/florence-2

and here's my Google Colab if you want to follow along

link: colab.research.google.com/github/roboflo…

Share this Scrolly Tale with your friends.

A Scrolly Tale is a new way to read Twitter threads with a more visually immersive experience.
Discover more beautiful Scrolly Tales like this.

Keep scrolling