[one-cmds] Introduce data post processing option to `one-infer`

[yunjayh]

WHAT

Let's introduce post-process stage by option in one-infer.

WHY

To support value-test between two different one-infer results, the inference result data should be well formatted. The standard format of data was decided to use h5 because h5 is a type which can contain multiple data with specific hierarchy and some other reasons. (There is a discussion in this link)

On the other hand, one-infer cannot guarantee various backend drivers emits well defined format data.

So, one-infer needs to solve that problem with post-process option which run data conversion after driver execution.

There is previous discussion about why post processing? in #9282.

HOW

Brief example usage will be like this.

$ ls
model.npu aaa.0.npy aaa.1.npy

# Note that [commands] run model.npu with aaa.0.npy and aaa.1.npy as input data 
# and emit output data as bbb.0.npy ...
# ./npy2h5.py can be any script which converts the npu-infer's result 
# to preset standard format of h5
$ one-infer -d npu-infer --post-process ./npy2h5.py -- [commands]
Inference complete!

$ ls
model.npu aaa.0.npy aaa.1.npy bbb.0.npy bbb.1.npy ccc.h5

IMHO --post-process ./npy2h5.py can be replaced to --post-process "./npy2h5.py aaa.0.npy aaa.1.npy bbb.0.npy ccc.h5" depends on how npy2h5.py has been implemented.

Any questions and suggestions are gratefully welcome!

[jinevening]

The idea looks good to me. Could you show the example of npy2h5.py? I'm curious about the format of the post-process script.

[yunjayh]

The idea looks good to me. Could you show the example of npy2h5.py? I'm curious about the format of the post-process script.

Of course. When I'm ready, I'll make and add it to my draft and ping you. It will be great to having one sample post-process script. :smile: Thank you!