I've added the Inception v3 model to our model collection:https://github.com/fchollet/deep-learning-models …
-
for example, the keras resnet has no weight reg, when the FB team trained it with 0.0001 l2 reg on bias + weights.
L2/L1 regularization is basically always harmful on this type of model. Regularization should be done via dropout/noise/etc.
-
-
The Facebook AI team trained resnets with l2 reg, and so did the original MS research paper:https://github.com/facebook/fb.resnet.torch …
-
I've not done grid search on the optimal way to regularize - just looked closely at how the original teams handled it though.
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.