bremen79 / cocob Goto Github PK

great code thanks

but do you some link to just plain python optimizer build on coin betting for usual tabular data?

It would be great if this could be installed through pypi.python.org

Whether your code var_update_op = state_ops.assign(var, var_update) change the w1,i of COCOB-Backprop Algorithm?

In your paper Training Deep Networks without Learning Rates
Through Coin Betting
You said that:
Note that the update in line 10 is carefully defined: The algorithm does not use the previous wt,i in
the update. Indeed, this algorithm belongs to the family of the Dual Averaging algorithms, where the
iterate is a function of the average of the past gradients [Nesterov, 2009].

You do so in Algorithm 2 COCOB-Backprop line 10

I was just wondering if you will try to add this to the official TF repo?

It will be extremely helpful if this can work in Keras.