I’ve discovered what Roboflow had to offer a few days ago and started reading about it and testing some features with the free plan, and it seems like a really powerful tool for computer vision and MLOps!
I am considering presenting Roboflow to my organisation’s board and suggesting moving out to it. Our primary use-case for CV is instance segmentation. Although YOLO seems to be performing well (on Coco), I am a bit afraid of switching to Roboflow if it means we’ll be locked to specific architectures. I believe the only current model options for segmentation are YOLO 11 and Roboflow 3.0.
So I just wanted to post here to see if I was missing anything. Is it somehow possible to create our own model architecture (E.G: using Pytorch), and import it in Roboflow and leverage the datasets versioning and model evaluation tools? If not, is it a feature that is considered by Roboflow for the future?
I also have a question about uploading models with customs weights: how easy is it to version these models along with the models trained on Roboflow and the datasets we version here? Is the model evaluation and all the metrics calculation handled by Roboflow automatically after uploading the custom weights? In other words, how can one assess the performance of multiple local experiments (models trained locally, and uploaded as weights) with the models trained on the Roboflow platform?
Thank you very much for your help, looking forward to reading your answers and discussing this
We’re always on the lookout for new models, though we’ve opted to curate a list of the best available models that we know we can support well across both edge and cloud deployment methods. For instance, we’ve recently added support for RF-DETR and Florence-2.
You can see the model architectures we support for upload here. Once uploaded, you’ll have model evaluation available (on paid plans).
We don’t currently support custom weights architectures - does that make up the majority of the models you have in production?
I just watched your live stream about RF-DETR, and it sounds very promising!
It’s also great that we can get model evaluations for custom weights To be honest the more I learn about Roboflow, the more impressed I am
We don’t currently support custom weights architectures - does that make up the majority of the models you have in production?
The majority of our models use common architectures and are segmentation models, but we have a few custom architectures as well. I see moving to Roboflow as moving to an End-To-End AI solution, for the long-term. So ideally, I wouldn’t want to maintain separate model architectures in separate repositories, nor to be locked to certain architectures forever
What would be your vision for the future of Roboflow? Could support for custom weights architectures ever be a supported feature?
The vision is that we’re a developer tool that can handle the end-to-end model lifecycle (annotate, train, evaluate, deploy, improve). This means we’re easy to integrate with + you can pick and choose which pieces of the platform you want to use.
We definitely want to be able to support custom architectures in the future. For now, we’re finding the best open source models and investing in those (as well as training our own!)
