Discrepancy between python and go outputs about tfgo HOT 2 CLOSED

Hi, I'm pretty sure the problem is in the preprocessing function. The resize functions are implemented differently from numpy/pillow/tensorflow/sklern/pytorch and thus the input values are different.

Read this article: https://zuru.tech/blog/the-dangers-behind-image-resizing

I suggest you move the pre-processing step inside the TensorFlow graph: in this way, your model will pre-process the input using only pure TensorFlow operations and the behavior will be consistent between every execution environment

from tfgo.

Thank you for your suggestion. I have taken care to ensure that the pre-processing steps in both Python and Golang produce the same output values, albeit with a different order of dimensions. Specifically, in Python, the input has a shape of (1,3,262,192), whereas in Golang, it has a shape of (1,192,262,3).

However, when attempting to reshape the Golang input to match the Python input, an error message is generated, indicating that the number of groups in the convolution layer is unevenly divisible by the output depth. This error message could be related to the way the input is being processed or the way the convolution layer is defined in the model. Here is the error:

panic: output depth must be evenly divisible by number of groups: 10 vs 64
	 [[{{function_node __inference___call___148}}{{node convolution}}]]

Interestingly this shape mismatch also exists with arrays of ones or zeros. But the model's outputs are the same in those cases. So that's why this discrepancy is confusing to me.

from tfgo.