NovoGradStep¶
The Novograd optimizer step.
Based on Ginsburg et al, 2019, Novograd is more robust to initial learning rate and weight initialization than other optimizers, and can for instance be used without learning-rate warm-up. The optimizer also works very well for large batch sizes, and has shown to be effective for up to 32k exemplars. Like all step nodes, this node only processes gradients, and the resulting updates must be applied manually to the weights (this can be accomplished using the Add node). However, you can also pass it to the StepSolver node which implements the full optimization loop. The learning rate can instead be given as a schedule, by wiring one of the Schedule nodes into the learning_rate_schedule port. More Info... Version 0.2.0
Ports/Properties¶
gradients¶
Gradients to be transformed.
weights¶
Optional current weights.
state¶
Explicit state of the node.
learning_rate_schedule¶
Optional learning rate schedule.
learning_rate¶
Learning rate. A typical choice may be 0.001 here, but this is problem dependent. If a learning rate schedule is provided, this value should be left unspecified.
beta1¶
Exponential decay rate for the first moment estimates.
beta2¶
Exponential decay rate for the second moment estimates.
epsilon¶
Small value applied to the denominator outside the square root to avoid dividing by zero when rescaling.
epsilon_inroot¶
Small value applied to the denominator inside the square root to avoid dividing by zero when rescaling. A case where this is needed is when differentiating the optimizer itself, eg for bilevel optimization.
weight_decay¶
Strength of the weight decay. This is multiplied by the learning rate as in e.g., PyTorch and Optax, but differs from the paper, where it is only multiplied by the schedule multiplier but not the base learning rate.
set_breakpoint¶
Set a breakpoint on this node. If this is enabled, your debugger (if one is attached) will trigger a breakpoint.
metadata¶
User-definable meta-data associated with the node. Usually reserved for technical purposes.