My experience so far

• Youtube videos on roboflow are awesome and explain a lot about ML and the latest trends, but actually trying it out is a pretty terrible experience (see below)
• Just casually browsing the universe leads to 429 Too many requests
• Cloning images from a universe project into my own project doesn’t show a loading spinner, and leaves me wondering what happened. After a while I realised there’s a cron job or something cloning images from Universe into my project, but it’s slow and I seem to have to refresh the page to see what’s happening even though it’s only 100 images.
• No way to delete images from my dataset without manually going into each image detail page and clicking the delete button there. I have to delete the entire project and start over
• Searching for “roboflow delete image” on Google, then clicking on the KB article leads to a 404: How do I delete a dataset? - Roboflow
• On Roboflow Universe, why is there no “Add dataset to your project” button?
• When cloning images, I can select a range of (max 250) images, but what if it’s a dataset of 1000 images? Browse over 40 pages and select all of them every time, hoping the cron job doesn’t choke?
• Training Failed. “Please try again with a new version and different export settings” is very vague and annoying. What was the real reason training failed? You expect me to randomly try out new options until it works without having any insights into what’s happening?
At this point I’m pretty ready to throw the towel in the ring and start training manually

Hi @nikke - very sorry to hear you’ve faced this challenges – my comments below are so we can glean more information to remedy these issues. And provide a better experience.

These issues with Dataset Search and Cloning Images from Universe – is there:

  1. a specific prompt, or prompts, you used where you saw the 429 errors?
  2. what dataset did you attempt to clone that failed?

The 2 questions are to make sure I understand the nature of the issue, and can attempt reproduction of it, myself – this will help us to get this all remedied.

For this issue – did you want to delete an entire batch, or just a select few images? We do have batch deletion if you needed the former:

The “Add dataset to your project” option – good call! That is something we do / are looking at, but I do want to add that if there is a dataset you wish to clone or add to your project, there is a Clone Images option (I can see why you don’t trust it and asked about this other solution due to the issues experiences thus far, so I want to be sure we assuage this issue, regardless of new product additions)

Yes, I can see how/why this is very frustrating. Is there a raw version of the dataset that was generated in the project? Raw versions are those without any preprocessing or augmentation features included, or those with only Auto-Orient for preprocessing and no augmentations.

  • If there is not a raw version, happy to generate one so you can use it in the meantime
  • If there is a raw version, I’d recommend exporting the raw version in COCO JSON, but if it is Classification then export in one Folder Structure or OpenAI CLIP Structure.
  • Its hard to say exactly, but I would like to see how we can improve the error message. Can you tell me more about your project?
  1. how many images were in the generated version?
  2. are you detecting small objects? or using aerial imagery?
  3. what is your class balance in the dataset like?
  4. what is the project type you trained? (e.g object detection, instance segmentation, etc)

We offer AutoML, and manual training, either works with Roboflow. Once again, I’d like to ensure we do resolve existing bugs, to improve the system, and also provide feedback to my team for feature considerations. If you do prefer manual training, we have over 30 notebooks available you can use in our notebooks repository, and model library:

  1. GitHub - roboflow/notebooks: Examples and tutorials on using SOTA computer vision models and techniques. Learn everything from old-school ResNet, through YOLO and object-detection transformers like DETR, to the latest models like Grounding DINO and SAM.
  2. Explore Top Computer Vision Models

^ above finding the information in the above image – use this docs page to start:

Thanks for highlighting a number of specific ways we can improve the Universe experience. @Mohamed did a nice job identifying specific ways to e.g. cloning images and training your own models for deployment.

Broadly, I agree with your feedback for areas we can make the open source dataset --> hosted app experience simpler.