Hello,
I know that axelera does not work in the attention mechanism. For example, can we convert the attention mechanism in the yolov11 architecture to convolution? Have you done any research or read on this subject?
Does Softmax structure work in Axelera? (as far as I know, no) Is there a way to turn Softmax into a working structure?
Hi Oguz
Interesting idea! This could be something really interesting to experiment with. Indeed, previous YOLO models have used pure-convolution architectures to achieve great results.
If you want to experiment building models like this, you can find our list of supported operators here: https://github.com/axelera-ai-hub/voyager-sdk/blob/59c4894b1bb4e7b17a6da92223bec5fdd2421fee/docs/reference/onnx-opset14-support.md#L4
With this, you can make sure that any customizations you make are compatible
Hi
Good news, the team is working on Yolov11 support, including the attention mechanism.
Stay tuned!
Free
voyager // wingman
Describe your vision pipeline in plain words. Wingman writes the YAML, runs it live on your Axelera hardware, and irons out the bugs with you.
Launch Wingman → Register for a free account
10 kits to win
the prompt // now open
Pitch it for a chance to win one of 10 Dell XE5s + Axelera Edge 130p to build it on. Closes soon.
Enter The Prompt → Register to enterAlready have an account? Login
Enter your E-mail address. We'll send you an e-mail with instructions to reset your password.
