Different version with same h5 file get different Evaluating?

[titu1994]

Different version with same h5 file get different Evaluating?

[titu1994]

Hmm thats definitely weird. The code diff basically shouldnt account for the change in scores. Can you try with the MobileNet snapshot weights from 0.3 and see the results?

Also, can you try 0.5 with imagesize of 224x224 ? I think size of the input image definitely can affect the result, but to completely change the ranking of a bad image to good is extremely weird.

[butterl]

tried with python3 evaluate_mobilenet.py -dir ./inputs -resize false and got the the same outputs as 0.5 version， seems the problem is here

-resize true will only cut image into 224 * 224 pieces then just evaluate the first 224 * 224 piece?

[titu1994]

Not cut. It will resize the whole image to that size and then score.

[butterl]

so the problem is that PIL resize image which make information lost？Do you got any good idea how to evaluate with big image ？ resize seem not work very well

[titu1994]

No idea. I guess you could take multiple crops of the same image and then average the scores but that defeats the purpose of a fast inference model.

[tech-tube-code]

@titu1994
Is it possible to split out evulation_mobilenet.py into 2 functions. One function will load the function and the other function will predict the score. Actually, I'm developing a flask app where I want to load all models when starting the flask app and through API just get the image score.