tensorflow: TFLite GPU Delegate has problem with MobileNetV2

System information

Have I written custom code (as opposed to using a stock example script provided in TensorFlow): Yes
OS Platform and Distribution : Linux Ubuntu 16.04
Mobile device : Samsung Galaxy S9 Android 8.0 - Nexus 10 Android 5.1
TensorFlow installed from (source or binary): Binary
TensorFlow version (use command below): b’v1.12.0-rc2-3-ga6d8ffae09’ 1.12.0 (0.0.0-gpu-experimental for mobile device)
Python version: 3.6

Describe the current behavior GPU delegate has problem with MobileNetV2. When I select GPU from device list at tflite demo project (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/java/demo) the app crashes. The only things I’ve changed on this project was changing MobileNet V1 float model to MobileNet v2. The MobileNetV2 model is taken from “https://tfhub.dev/google/imagenet/mobilenet_v2_050_160/classification/2” , retrained by “https://github.com/tensorflow/hub/blob/master/examples/image_retraining/retrain.py” script and converted to tflite format using the following command:

tflite_convert \
  --output_file=graph.tflite \
  --graph_def_file=retrained_graph.pb \
  --input_arrays=Placeholder \
  --output_arrays=final_result
 --input_shapes=1,160,160,3

All the necessary changes (such as changing the graph name, input size, etc.) in the ImageClassifierFloatMobileNet class is made.

Logs:

2019-01-23 13:51:12.091 22222-22294/android.example.com.tflitecamerademo E/AndroidRuntime: FATAL EXCEPTION: CameraBackground
    Process: android.example.com.tflitecamerademo, PID: 22222
    java.lang.IllegalArgumentException: Internal error: Failed to apply delegate: GpuDelegate Prepare: Dimension is empty.Node number 68 (GpuDelegate) failed to prepare.
    
        at org.tensorflow.lite.NativeInterpreterWrapper.applyDelegate(Native Method)
        at org.tensorflow.lite.NativeInterpreterWrapper.init(NativeInterpreterWrapper.java:83)
        at org.tensorflow.lite.NativeInterpreterWrapper.<init>(NativeInterpreterWrapper.java:60)
        at org.tensorflow.lite.Interpreter.<init>(Interpreter.java:224)
        at com.example.android.tflitecamerademo.ImageClassifier.recreateInterpreter(ImageClassifier.java:168)
        at com.example.android.tflitecamerademo.ImageClassifier.useGpu(ImageClassifier.java:176)
        at com.example.android.tflitecamerademo.Camera2BasicFragment.lambda$updateActiveModel$0$Camera2BasicFragment(Camera2BasicFragment.java:379)
        at com.example.android.tflitecamerademo.Camera2BasicFragment$$Lambda$0.run(Unknown Source:8)
        at android.os.Handler.handleCallback(Handler.java:789)
        at android.os.Handler.dispatchMessage(Handler.java:98)
        at android.os.Looper.loop(Looper.java:164)
        at android.os.HandlerThread.run(HandlerThread.java:65)
2019-01-23 13:51:12.102 4411-8924/? E/CameraDeviceClient: Disconnect from CameraDeviceClient

About this issue

Original URL
State: closed
Created 5 years ago
Comments: 16 (10 by maintainers)

Most upvoted comments

@aselle Thanks! 👍

@ramtin2080

The GPU delegate currently does not support dynamic tensor sizes. Is it possible for you to explicitly specify the tensor dimensions of module_apply_default/hub_input/Mul/y and module_apply_default/hub_input/Sub/y in your TF graph, so that all tensor dimensions are known in advance?

impjdi on Jan 30, 2019

@ramtin2080

I just noticed:

GPU delegate has problem with MobileNetV2. When I select GPU from device list at tflite demo project (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/java/demo) the app crashes.

Let me try to take a look at these constants. Maybe we can apply that in the Java layer directly. Stay tuned.

impjdi on Feb 1, 2019

@ramtin2080

Aite. I looked into the constants, and it looks like it’s doing 2.0 * X - 1.0 which is remapping the values from 0.0 - 1.0 to -1.0 - 1.0. So let’s do that.

imgData.putFloat(((pixelValue >> 16) & 0xFF) / 255.f * 2.f - 1.f); imgData.putFloat(((pixelValue >> 8) & 0xFF) / 255.f * 2.f - 1.f); imgData.putFloat((pixelValue & 0xFF) / 255.f * 2.f - 1.f);

Again, that’s the code in ImageClassifierFloatMobileNet.java.

impjdi on Feb 1, 2019