java - 在 resizeInput() 后运行时，Tensorflow lite 解释器崩溃

Question

我有一个使用的 tensorflow lite 模型，输入是{1,320,320,3} 输出数组数据类型是 {FLOAT32,FLOAT32,INT32}。将输入大小调整为{1,224,320,3}时，Interpreter#runForMultipleInputsOutputs 崩溃。

我的张量流精简版：

    implementation('org.tensorflow:tensorflow-lite:0.0.0-nightly') { changing = true }
    implementation('org.tensorflow:tensorflow-lite-gpu:0.0.0-nightly') { changing = true }
    implementation('org.tensorflow:tensorflow-lite-support:0.0.0-nightly') { changing = true }

创建输入张量。准备输出张量：

        inputImageBuffer = new TensorImage(imageDataType);
        outputArr = new TensorBuffer[tflite.getOutputTensorCount()];
        resizeOutputArr = new TensorBuffer[tflite.getOutputTensorCount()];
        for (int i = 0, count = tflite.getOutputTensorCount(); i < count; i++) {
            DataType pDataType = tflite.getOutputTensor(i).dataType();
            if (pDataType == DataType.INT32) pDataType = DataType.FLOAT32;
            outputArr[i] = TensorBuffer.createFixedSize(i == 0 ? new int[]{imageSizeY, imageSizeX} : new int[]{imageSizeX}, pDataType);
            resizeOutputArr[i] = TensorBuffer.createDynamic(pDataType);
        }

调整输入图像缓冲区的大小

        // Loads bitmap into a TensorImage.
        inputImageBuffer.load(bitmap);
        float sc = bitmap.getWidth() / 320.0f;
        int scaleWid = Math.round(bitmap.getWidth() / sc);
        int scaleHei = Math.round(bitmap.getHeight() / sc);
        imageSizeY = scaleHei / 32 * 32;
        // Creates processor for the TensorImage.
        int numRotation = sensorOrientation / 90;
        this.processor =
                new ImageProcessor.Builder()
                        .add(new ResizeOp(scaleHei, scaleWid, ResizeOp.ResizeMethod.NEAREST_NEIGHBOR))
                        .add(new ResizeWithCropOrPadOp(imageSizeY, imageSizeX))
                        .add(new Rot90Op(numRotation))
                        .add(new GrayOp(imageSizeY, imageSizeX))
                        .add(getPreprocessNormalizeOp())
                        .build();
        processor.process(inputImageBuffer);

跑：

        int imageTensorIndex = 0;
        int[] imageShape = tflite.getInputTensor(imageTensorIndex).shape(); // {1, height, width, 3}
        if (imageSizeY != imageShape[1] || imageShape[1] == 0) { // imageSizeY changed
            tflite.resizeInput(imageTensorIndex, new int[]{1, imageSizeY, imageSizeX, 3}); //resize inputTensor
            for (int i = 0, count = tflite.getOutputTensorCount(); i < count; i++) { //recreate outputTensor
                DataType pDataType = tflite.getOutputTensor(i).dataType();
                if (pDataType == DataType.INT32) pDataType = DataType.FLOAT32;
                outputArr[i] = TensorBuffer.createFixedSize(i == 0 ? new int[]{imageSizeY, imageSizeX} : new int[]{imageSizeX}, pDataType);
            }
        }
        Map<Integer, Object> outputMap = new HashMap<>();
        for (int i = 0, cnt = outputArr.length; i < cnt; i++) { //rewind outputBuffer
            outputMap.put(i, outputArr[i].getBuffer().rewind());
            resizeOutputArr[i].getBuffer().rewind();
        }
        try {
            tflite.runForMultipleInputsOutputs(new Object[]{inputImageBuffer.getBuffer()}, outputMap);
        } catch (IllegalArgumentException e) {
            e.printStackTrace();
            return null;
        }
        for (int i = 0, count = tflite.getOutputTensorCount(); i < count; i++) { //resize outputArr with new shape
            int[] shape = tflite.getOutputTensor(i).shape();
            outputArr[i].getBuffer().limit(computeFlatSize(shape) * DataType.FLOAT32.byteSize());
            resizeOutputArr[i].loadBuffer(outputArr[i].getBuffer(), shape);
            outputArr[i].getBuffer().clear();
            if (i == count - 1 && shape != null && shape[0] <= 0) return null;
        }
        return resizeOutputArr;

    protected static int computeFlatSize(@NonNull int[] shape) {
        SupportPreconditions.checkNotNull(shape, "Shape cannot be null.");
        int prod = 1;
        for (int i = 0; i < shape.length; ++i) {
            prod *= shape[i];
        }
        return prod;
    }

当位图中没有可识别的对象时，我得到了异常：

W: java.lang.IllegalArgumentException: Internal error: Tensor hasn't been allocated.
W:     at org.tensorflow.lite.Tensor.buffer(Native Method)
W:     at org.tensorflow.lite.Tensor.buffer(Tensor.java:493)
W:     at org.tensorflow.lite.Tensor.copyTo(Tensor.java:264)
W:     at org.tensorflow.lite.Tensor.copyTo(Tensor.java:254)
W:     at org.tensorflow.lite.NativeInterpreterWrapper.run(NativeInterpreterWrapper.java:170)
W:     at org.tensorflow.lite.Interpreter.runForMultipleInputsOutputs(Interpreter.java:343)
W:     at com.fotric.irdetector.qf.tflite.Classifier.recognizeImage(Classifier.java:283)

这个异常会降低识别性能，有人帮帮我，谢谢！！！

标签： javaandroidtensorflow-lite