@Operator(group="train") public final class ResourceSparseApplyFtrl extends PrimitiveOp
That is for rows we have grad for, we update var, accum and linear as follows: grad_with_shrinkage = grad + 2 * l2_shrinkage * var accum_new = accum + grad_with_shrinkage * grad_with_shrinkage linear += grad_with_shrinkage + (accum_new^(-lr_power) - accum^(-lr_power)) / lr * var quadratic = 1.0 / (accum_new^(lr_power) * lr) + 2 * l2 var = (sign(linear) * l1 - linear) / quadratic if |linear| > l1 else 0.0 accum = accum_new
Modifier and Type | Class and Description |
---|---|
static class |
ResourceSparseApplyFtrl.Options
Optional attributes for
ResourceSparseApplyFtrl |
operation
Modifier and Type | Method and Description |
---|---|
static <T,U extends Number> |
create(Scope scope,
Operand<?> var,
Operand<?> accum,
Operand<?> linear,
Operand<T> grad,
Operand<U> indices,
Operand<T> lr,
Operand<T> l1,
Operand<T> l2,
Operand<T> l2Shrinkage,
Operand<T> lrPower,
ResourceSparseApplyFtrl.Options... options)
Factory method to create a class wrapping a new ResourceSparseApplyFtrl operation.
|
static ResourceSparseApplyFtrl.Options |
useLocking(Boolean useLocking) |
equals, hashCode, op, toString
public static <T,U extends Number> ResourceSparseApplyFtrl create(Scope scope, Operand<?> var, Operand<?> accum, Operand<?> linear, Operand<T> grad, Operand<U> indices, Operand<T> lr, Operand<T> l1, Operand<T> l2, Operand<T> l2Shrinkage, Operand<T> lrPower, ResourceSparseApplyFtrl.Options... options)
scope
- current scopevar
- Should be from a Variable().accum
- Should be from a Variable().linear
- Should be from a Variable().grad
- The gradient.indices
- A vector of indices into the first dimension of var and accum.lr
- Scaling factor. Must be a scalar.l1
- L1 regularization. Must be a scalar.l2
- L2 shrinkage regulariation. Must be a scalar.l2Shrinkage
- lrPower
- Scaling factor. Must be a scalar.options
- carries optional attributes valuespublic static ResourceSparseApplyFtrl.Options useLocking(Boolean useLocking)
useLocking
- If `True`, updating of the var and accum tensors will be protected
by a lock; otherwise the behavior is undefined, but may exhibit less
contention.Copyright © 2022. All rights reserved.