Change the inference example to an unittest

[Xreki]

fix #7876

[kexinzhao]

Why are the numpy array dimensions different for 'mlp' and 'conv_net' ?

[Xreki]

Because the input's dimension of conv_op should be 4-D or 5-D, and the input's dimension of mul_op will be flattened into 2-D. So, we can use line 169 to simplify the code.

[kexinzhao]

LGTM!

Maybe wait for @luotao1 and @sidgoyal78 to also have a look?

A separate question would be "are we going to merge this before #7690"?

[sidgoyal78]

Thanks for the PR. I have one suggestion: for adding the inference example in the python side, can we create another folder?
I remember the test_recognize_digits.py was modified significantly by someone few days back (and in that process the inference parts were simply deleted). So maybe just to avoid general conflict, we can do some code duplication in order to keep things independent?

[luotao1]

LGTM

[Xreki]

for adding the inference example in the python side, can we create another folder?

If doing this, there will be so many redundant codes. All the network will be write twice and train twice, because we need to use the training results.

I remember the test_recognize_digits.py was modified significantly by someone few days back (and in that process the inference parts were simply deleted).

It won't happen again, because I change the inference example to an unittest. If someone deletes the python codes again, the CI will fail.