How to set IMAGE_MEAN and IMAGE_STD

[googlecodelabs]

How to set IMAGE_MEAN and IMAGE_STD

[MarkDaoust]

Hi,

99.9 sounds a little high ... are you cross validating to make sure you aren't overfitting your data?

IMAGE_MEAN and IMAGE_STD aren't meant to be a function of the images. They're more a question of the expected input range of the the network being used. Some nets want the input on the [0,255] range, others [0,1], others [-1,1],... and having these "mean" and "variance" parameters to do what looks like a Z-score normalization (subtract the mean, and divide by the variance) is a simple way to enforce that.

I really don't think they're the cause of your problem.

Good luck!

[MarkDaoust]

If you really want to try, the values are right there in ClassifierActivity.java but you should not have to update these, I'm prrety sure it can only make things worse.

[Digital2Slave]

@MarkDaoust Thanks a lot. I just followed the steps post in tensorflow-for-poets-2. I will do cross validating to make sure whether the model overfitting or not.
The retrain.py is here which seem like contain validation sets.