Skip to content

← deep_learning package

LAMBStep

The LAMB optimizer step.

Based on You et al, 2019, LAMB is an optimizer that combines the versatility of Adam with ability to operate on both small and very large batch sizes (e.g., 16k examples) using the strategy employed by LARS, which is a precursor to LAMB. 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. The weight decay can be used in conjunction with a mask data structure that has the same nested structure as the weights being optimized, but which contains booleans indicating which weights should be decayed. More Info... Version 0.2.0

Ports/Properties

gradients

Gradients to be transformed.

verbose name
Gradients
default value
None
port type
DataPort
value type
object (can be None)
data direction
INOUT

weights

Optional current weights.

verbose name
Weights
default value
None
port type
DataPort
value type
object (can be None)
data direction
IN

state

Explicit state of the node.

verbose name
State
default value
None
port type
DataPort
value type
object (can be None)
data direction
INOUT

learning_rate_schedule

Optional learning rate schedule.

verbose name
Learning Rate Schedule
default value
None
port type
DataPort
value type
BaseNode (can be None)
data direction
IN

weight_decay_mask

Mask structure for the weight decay.

verbose name
Weight Decay Mask
default value
None
port type
DataPort
value type
object (can be None)
data direction
IN

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.

verbose name
Learning Rate
default value
None
port type
FloatPort
value type
float (can be None)

beta1

Exponential decay rate for the first moment estimates.

verbose name
Beta1
default value
0.9
port type
FloatPort
value type
float (can be None)

beta2

Exponential decay rate for the second moment estimates.

verbose name
Beta2
default value
0.999
port type
FloatPort
value type
float (can be None)

epsilon

Small value applied to the denominator outside the square root to avoid dividing by zero when rescaling.

verbose name
Epsilon
default value
1e-06
port type
FloatPort
value type
float (can be None)

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.

verbose name
Epsilon (Inside Root)
default value
0
port type
FloatPort
value type
float (can be None)

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.

verbose name
Optional Weight Decay
default value
0
port type
FloatPort
value type
float (can be None)

set_breakpoint

Set a breakpoint on this node. If this is enabled, your debugger (if one is attached) will trigger a breakpoint.

verbose name
Set Breakpoint (Debug Only)
default value
False
port type
BoolPort
value type
bool (can be None)

metadata

User-definable meta-data associated with the node. Usually reserved for technical purposes.

verbose name
Metadata
default value
{}
port type
DictPort
value type
dict (can be None)