tensorflow: Android "TF Stylize" demo crash with "input_max_range must be larger than input_min_range"

I have compiled and installed the demo app according to the instruction in the README. The “TF Detect” and “TF Classify” apps are working as expected. However, starting the “TF Stylize” app, it always crashes in about 3 seconds. The prebuilt apk crashes in the same way. Here is the android logcat recorded:

$ adb -s b80360c7 logcat | grep Tensor I/TensorFlowInferenceInterface( 6801): Checking to see if TensorFlow native methods are already loaded I/TensorFlowInferenceInterface( 6801): TensorFlow native methods not found, attempting to load via tensorflow_inference I/TensorFlowInferenceInterface( 6801): Successfully loaded TensorFlow native methods (RunStats error may be ignored) I/TensorFlowInferenceInterface( 6801): Model load took 540ms, TensorFlow version: 1.0.1 I/TensorFlowInferenceInterface( 6801): Successfully loaded model from ‘file:///android_asset/stylize_quantized.pb’ E/TensorFlowInferenceInterface( 6801): Failed to run TensorFlow inference with inputs:[input, style_num], outputs:[transformer/expand/conv3/conv/Sigmoid] E/TensorFlowInferenceInterface( 6801): Inference exception: java.lang.IllegalArgumentException: input_max_range must be larger than input_min_range. E/TensorFlowInferenceInterface( 6801): [[Node: transformer/contract/conv1/Relu_eightbit_quantize_transformer/contract/conv1/InstanceNorm/batchnorm/add_1 = QuantizeV2[T=DT_QUINT8, mode=“MIN_FIRST”, _device=“/job:localhost/replica:0/task:0/cpu:0”](transformer/contract/conv1/InstanceNorm/batchnorm/add_1, transformer/contract/conv1/Relu_eightbit_min_transformer/contract/conv1/InstanceNorm/batchnorm/add_1, transformer/contract/conv1/Relu_eightbit_max_transformer/contract/conv1/InstanceNorm/batchnorm/add_1)]] E/AndroidRuntime( 6801): at org.tensorflow.contrib.android.TensorFlowInferenceInterface.getTensor(TensorFlowInferenceInterface.java:486) E/AndroidRuntime( 6801): at org.tensorflow.contrib.android.TensorFlowInferenceInterface.readNodeIntoFloatBuffer(TensorFlowInferenceInterface.java:332) E/AndroidRuntime( 6801): at org.tensorflow.contrib.android.TensorFlowInferenceInterface.readNodeFloat(TensorFlowInferenceInterface.java:287)

As can be inferred from the logcat, the problem happens in tensorflow/core/kernels/quantize_op.cc line 71. The app is using a quantized model. I found that input_min_range is set to be inf, and input_max_range is -inf, which is obviously wrong.

What should be done to fix it?

The crash appears on Samsung Galaxy S4, but does not appears on emulator with Intel CPU.

[Edit] With more experiment I found that sometimes the apps works without crashing, about 1 out of 10 trials.

About this issue

Original URL
State: closed
Created 7 years ago
Comments: 49 (27 by maintainers)

Most upvoted comments

Ok, so 1920x1080 produces the inference error, while 1440x1080 complains about invalid size? That makes more sense.

Three possible options here when legacy mode is detected:

Only use the luminance plane to create the image. It would be something like:

byte[] inputBuffer = // from preview frame plane 0
int[] bitmapBuffer = new int[frameHeight * frameWidth];
for (int y = 0; y < frameHeight; ++y) {
  for (int x = 0; x < frameWidth; ++x) {
    int pixelValue = inputBuffer[y * frameWidth + x];
    bitmapBuffer[y * frameWidth + x] = Color.ARGB(255, value, value, value);
  }
}

Then bitmapBuffer would be used to initialize the RGB bitmap before it gets resized.

Use the android.hardware.Camera API rather than Camera2 and see if you have better luck. You could try updating the implementation here to work with the current demo code.
Figure out the correct conversion for whatever random color format your phone tries to use, but that would mean finding out programmatically what that is.

I’d suggest 1 or 2, which we would gladly welcome contributions for. 3 seems a bit too special-cased for this demo code, but if you just want to get it working for yourself is an option.

andrewharp on Mar 24, 2017