avg_mean and/or avg_var in BatchNormalization not working? about chainer HOT 3 CLOSED

Also, I should point out that when I try to try a decay value of 0.0, the network is kinda close in accuracy, but not really (it doesn't come that close to what would be expected given that the loss is so good during training; this probably isn't a case of overfitting as I'm testing, for now, on the same images that I've trained on).

Also, off topic, but I love this framework. It's a shame that PFN gave up on it. I find PyTorch to be spaghetti code.

from chainer.

Using the recommended 1.17 (well, 1.17.3) version of NumPy doesn't seem to help.

from chainer.

I wrote my own variant of Batch Normalization, and it appears to be having similar issues both evaluating on the network that is left in memory right after training as well as one loaded from disk (aka deserialized). I think this means that there isn't an issue with Chainer's BatchNormalization, but rather an issue with my network (or dataset).

from chainer.