reason behind low value of parameters in VGG19 about recipes HOT 1 CLOSED

Did your weights converge to this value during the course of training or did you use an external function to reduce the values?

The weights were provided by the original authors of the paper at http://www.robots.ox.ac.uk/%7Evgg/research/very_deep/. I'm sure they obtained these weights in training. The training protocol is described in http://arxiv.org/abs/1409.1556, Section 3.1.

I'm trying to train a VGG decoder (with Batch Normalization after every conv and TransposedConv layer in the decoder).

With batch normalization, the scale of weights does not matter anyway, it will be countered by normalizing to unit variance. You can add L2 penalties to the weights if you like smaller values.

Closing because it's a question, not anything we need to change in Lasagne Recipes. But feel free to continue the discussion here.

from recipes.