loss didn't get down

[chutongz]

Hi, I am using your code on both VGG and your original alexnet networks. The way I manage my data is exact following your instructions, and I tried to use np_util.to_categorial() function instead, but the loss is always around 8 even after 100+ epochs, and 16 on the original code. Do you have any idea what's the probable problem of it? Any help is appreciated, thanks a lot!

[matteo-dunnhofer]

Hi @chutongz have you already tried to lower the learning rate? Like to 1e-4?

[briansune]

@dontfollowmeimcrazy i do have the same situation with your posted code. If learning rate is set to 1e-2 the loss drop logically, while after 1 epoch the loss saturate and no longer drop. May i ask what is the final settings for a workable training?