替换自定义训练的TensorFlow Lite模型后Android应用崩溃求助

Let’s walk through the most likely causes and fixes for this crash—since your model works locally with label_image, the issue almost certainly lies in how the model is converted for TensorFlow Lite or how the Android app is configured to use it.

• Ensure You Properly Converted the .pb Model to .tflite
  The TensorFlow Lite Android app requires a .tflite file, not the raw .pb graph. If you skipped this step or did it incorrectly, that’s the first thing to fix. Use the tflite_convert tool with parameters matching your model’s specs:

  tflite_convert \
    --graph_def_file=tf_files/retrained_graph.pb \
    --output_file=tf_files/retrained_graph.tflite \
    --input_arrays=input \
    --output_arrays=final_result \
    --input_shapes=1,224,224,3
  

  • Double-check the input/output array names: You can verify these by loading the .pb model in TensorBoard or using saved_model_cli to inspect the graph.
  • Confirm the input shape matches your mobilenet_0.50_224 model (224x224x3).

• Verify Android App Input Preprocessing Matches Your Model
  MobileNet models expect specific input normalization, and mismatched preprocessing can cause memory access errors (like SIGBUS):

  • Open the ImageClassifier (or ClassifierQuantizedMobileNet) class in the Android demo.
  • Update these constants to match your model:
    • INPUT_SIZE = 224 (matches your model’s input dimension)
    • IMAGE_MEAN = 127.5
    • IMAGE_STD = 127.5 (these values normalize pixel data from 0-255 to [-1, 1], which MobileNet expects)
  • Ensure the number of output classes (OUTPUT_CLASSES) matches the number of flower categories in your training dataset.

• Check for Corrupted Model Files
  SIGBUS errors often happen when the app tries to read an incomplete or corrupted .tflite file:

  • Re-copy the converted .tflite file into your Android project’s assets directory.
  • Compare the file size of the .tflite file in your project with the one on your local machine to confirm they’re identical.

• Validate the TFLite Model Locally
  Test the converted .tflite model on your computer first to rule out conversion issues:

  import tensorflow as tf
  
  # Load the TFLite model
  interpreter = tf.lite.Interpreter(model_path="tf_files/retrained_graph.tflite")
  interpreter.allocate_tensors()
  
  # Check input/output specs
  input_details = interpreter.get_input_details()
  output_details = interpreter.get_output_details()
  
  print(f"Expected input shape: {input_details[0]['shape']}")  # Should be [1, 224, 224, 3]
  print(f"Expected input dtype: {input_details[0]['dtype']}")  # Typically float32 for unquantized MobileNet
  

  If this throws errors, your conversion step was faulty—go back to the first point and adjust the tflite_convert parameters.

• Match TensorFlow Versions
  Make sure the TensorFlow version used to train the model is compatible with the TensorFlow Lite version in your Android app:

  • If you trained with TensorFlow 1.x, use a 1.x-compatible TFLite dependency in your build.gradle (e.g., org.tensorflow:tensorflow-lite:1.15.0).
  • If using TensorFlow 2.x, ensure you used the correct conversion workflow (e.g., using tf.lite.TFLiteConverter instead of the old tflite_convert tool).

Start with the first two checks—these are the most common culprits when moving a working local model to Android. If the crash persists, share more details about your conversion command and Android code adjustments, and we can dig deeper.

内容的提问来源于stack exchange，提问作者Mandar