How to interpret the result returned by a vmfb file?

[FlintWangacc]

I encounter some issues when running vmfb file produced by iree.
I use the following steps to utilize the vmfb file.
First I dump the input tensor to disk through numpy

import numpy as np
...
np.save('input.npy', input_batch, allow_pickle=True)

Then I pass the input tensor to the neural network through iree-run-module

iree-run-module --module=file.vmfb --function=main --input="1x3x224x224xf32=@./input.npy" --output="@./output.npy

Then I pass the output.npy back to python to get the result.

import numpy as np
output = np.load('output.npy')
...

The output tensor has correct shape. But can't get correct result. Do you know how to fix it?