Custom Vision on HoloLens

2019-08-25 03:03发布

I'm using custom Vision ( https://www.customvision.ai ) to train a model for object recognition. After 10 Iterations of training it suddenly stopped loading. I always export it as onnx and load it on the HoloLens (with this tutorial: https://mtaulty.com/2018/03/29/third-experiment-with-image-classification-on-windows-ml-from-uwp-on-hololens-in-unity/ ). And it worked for quite some time (though the results weren't perfect), but after I continued to train the model to find the things better it simply stopped working.

My Scripts are pretty much like the ones in the Tutorial, just with my Tags and some different names. MainScript.cs https://pastebin.com/hxLrjXqv - my detect.cs (like the daschund.cs in the tutorial) https://pastebin.com/sP4PWNV6

So I deleted my project on Custom Vision and started trained my model from the ground up again. Only have like 40 pictures for 2 Tags with Multiclass (Single tag per image). I only want to know if there is a pointing finger in the picture or not. When deploying on the HoloLens it doesn't get over the line in MainScript

LearningModelEvaluationResultPreview evalResult = await learningModel.EvaluateAsync(binding, string.Empty);

I don't know what's the problem. If anyone has an idea how to fix it or a different source for how to learn how to run an Object Recognition Model on the HoloLens I would be happy.

Thanks in advance

2条回答
贪生不怕死
2楼-- · 2019-08-25 03:12

Custom Vision service updated to allow download of ONNX v1.2 models. Solved my issue.

查看更多
Explosion°爆炸
3楼-- · 2019-08-25 03:37

It could be due to the subscription level you are using. You subscription will limit the number of models you can have in storage as a result of training. You subscription is probably limiting you to retaining 10 iterations at a time.

If you go into the CustomVision.AI portal and look at your project, click on Performance tab and you will see all of your training iterations. Delete the oldest one and I suspect you will be able to train your model again, if I understand your problem correctly.

查看更多
登录 后发表回答