Anyone here experienced with tensorflow?
I'm thinking of some image recognition with tensorflow on mobile. Now the question is how performant will it be on older devices? Do I run on mobile with lite or do I run on server so I can support most devices or a hybrid of both? Anyone experienced this and how did you approach your decision?
I suggest move to ONNX if you can, opens up a wider range of deployment / device options.
How do I run ONNX on a mobile device?
ONNX Runtime is the most mature framework that has the most bindings/compute engines for different platforms (including mobile).
But there are also other options check https://onnx.ai for all supporters
I don't see the link for mobile. Where's the link for android/ios / hybrid? What's your experience with it and performance on older devices?
For iOS and Android they are still in preview so you would have to check the ONNXRuntime github page, as there is no distribution package for that one yet.
Yeah, I looked into it a bit more. I'm going to pass and stick to tensorflow. The only thing else I saw that looked a bit interesting for running on older devices is ncnn.
Sure that's perfectly fine if you don't mind being stuck in the google ecosystem. In my experience at least last time I tried ncnn was pretty crappy in terms of model support. But then again no idea what type of models you are trying to run. Also I would be very very careful with model pruning, looks good on paper but in reality it cuts off the long tail.
Thanks for all the suggests. I'm still learning. This is my first go round with it. I can always switch to a different frame work down the line. I'm afraid of being stuck in google ecosystem, which is what ONNX looks like it would be great for.
even old devices can run quite a few models, in the end it all comes to your model size. to calculate precisely please elaborate what you plan on doing exactly.
Do you have experience running on mobile? What was the spec of the old device you ran on? Did you run lite or js?
not too old, I ran mediapipe (js) on mobile but again it all depends on how many parameters you run
I'll like to support as many devices as possible. Training for image recognition. I tried the js, not bad on a modern device, horrible on older devices. Looks like I'm going to go native.