Cannot find 'MetalDelegate' in scope about tflite-support HOT 11 CLOSED

@khanhlvg do you have any idea why the MetalDelegate cannot be found?

from tflite-support.

@GeneralJing what was the error that you saw?

from tflite-support.

delegates = [MetalDelegate()] this line， ‘MetalDelegate' cannot find. MetalDelegate()is undefined.

from tflite-support.

@GeneralJing as per the tflite documentation:

"From 2.4.0 version and latest nightly releases, by default GPU and Core ML delegates are excluded from the pod to reduce the binary size. You can include them by specifying subspec:
pod 'TensorFlowLiteSwift', '~> 0.0.1-nightly', :subspecs => ['CoreML', 'Metal']"

So you'll have to add 'Metal' to your podfile and only then will it be available.

from tflite-support.

Thanks. I will try that later. Whatever result i have, i will leave a message here.

from tflite-support.

I tried it and it worked. Thank you. But the time is very long on iphone 6s.
Preprocessing:197ms
Model inference:56ms
Postprocessing:1591ms
Visualization:8ms.
How did you solve the time-consuming problem of tflite in practice？

from tflite-support.

Yeah I found it to be slow too. Didn't think it could be an issue with tflite implementation. I thought it must be related to model's size or something. I moved on from the model and am now looking for other alternatives.

from tflite-support.

Ok, thanks for your help.

from tflite-support.

Hey @GeneralJing, just found this note in the tflite sample codes:

// Note: You may find postprocessing very slow if you run the sample app with Debug build.
// You will see significant speed up if you rerun using Release build, or change
// Optimization Level in the project's Build Settings to the same value with Release build.

Also checkout the following:
https://blog.tensorflow.org/2020/07/accelerating-tensorflow-lite-xnnpack-integration.html

from tflite-support.

I only tested the relevant time on my mobile phone and did not notice the relevant notes. Thank you for your carefulness and reminder. The added link is very valuable, and I will study it carefully. If I encounter other related problems later, can I communicate with you by email？

from tflite-support.

Haha yeah, same with me. Also, one more thing I noticed was that the post-processing codes were not optimised (atleast for my use-case) and do not utilise the gpu/multi-threading either. So that is also something you can look into if you want improved performance.

Yeah sure, feel free to reach out.

from tflite-support.