Much like Adam is essentially RMSprop with momentum, Nadam is Adam RMSprop with Nesterov momentum.
optimizer_nadam(
learning_rate = 0.002,
beta_1 = 0.9,
beta_2 = 0.999,
epsilon = NULL,
schedule_decay = 0.004,
clipnorm = NULL,
clipvalue = NULL,
...
)float >= 0. Learning rate.
The exponential decay rate for the 1st moment estimates. float, 0 < beta < 1. Generally close to 1.
The exponential decay rate for the 2nd moment estimates. float, 0 < beta < 1. Generally close to 1.
float >= 0. Fuzz factor. If `NULL`, defaults to `k_epsilon()`.
Schedule deacy.
Gradients will be clipped when their L2 norm exceeds this value.
Gradients will be clipped when their absolute value exceeds this value.
Unused, present only for backwards compatability
Default parameters follow those provided in the paper.
[On the importance of initialization and momentum in deep learning](https://www.cs.toronto.edu/~fritz/absps/momentum.pdf).
Other optimizers:
optimizer_adadelta(),
optimizer_adagrad(),
optimizer_adamax(),
optimizer_adam(),
optimizer_rmsprop(),
optimizer_sgd()