Java中的Tensorflow：inferenceInterface.fetch转换为多维数组

Question

I am using a trained Tensorflow model in Java on Android. I am trying to extract the output of an intermediary op.

The Tensor I am extracting has shape (150, 150, 256).

I have declared the output target to be

private float[] hybridValues;
hybridValues = new float[150 * 150 * 256];

I am then getting the output using.

inferenceInterface.fetch(OUTPUT_NODE, hybridValues);

The values are fine but they are stored as a 1D array. Is there a way to get inferenceinterface.fetch to return a multi-dimentional array?

I tried declaring hybridValue to be a three dimensional float array but that doesn't work as the fetch method expects a 1D array.

The end goal is to pass my output to a Python program that will feed the values to a Tensor of the same shape (150, 150, 256).

For comparison the Python a_output = graph.get_tensor_by_name('a2b_generator/Conv_7/Relu:0') returns an ndarray with values in the same shape as the target tensor.

Answer 1

我在桌面上使用tensorflow和java（可能有所不同），我要做的就是创建一个具有正确大小的多维数组，然后将值复制到上面，例如Tensor.copyTo(Object dst)