org.apache.sysml.api.dl.CaffeSolver.scala Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of systemml Show documentation
Show all versions of systemml Show documentation
Declarative Machine Learning
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing,
* software distributed under the License is distributed on an
* "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
* KIND, either express or implied. See the License for the
* specific language governing permissions and limitations
* under the License.
*/
package org.apache.sysml.api.dl
import caffe.Caffe.SolverParameter
import org.apache.sysml.runtime.DMLRuntimeException
import caffe.Caffe
trait CaffeSolver {
def sourceFileName: String;
def update(dmlScript: StringBuilder, layer: CaffeLayer): Unit;
def init(dmlScript: StringBuilder, layer: CaffeLayer): Unit;
// ----------------------------------------------------------------
// Used for Fine-tuning
private def getLayerLr(layer: CaffeLayer, paramIndex: Int): String = {
val param = layer.param.getParamList
if (param == null || param.size() <= paramIndex)
return "lr"
else
// TODO: Ignoring param.get(index).getDecayMult for now
return "(lr * " + param.get(paramIndex).getLrMult + ")"
}
// the first param { } is for the weights and the second is for the biases.
def getWeightLr(layer: CaffeLayer): String = getLayerLr(layer, 0)
def getBiasLr(layer: CaffeLayer): String = getLayerLr(layer, 1)
// ----------------------------------------------------------------
def commaSep(arr: String*): String =
if (arr.length == 1) arr(0)
else {
var ret = arr(0)
for (i <- 1 until arr.length) {
ret = ret + "," + arr(i)
}
ret
}
def regularization_update(regularizationType:String, lambda: Double, dmlScript: StringBuilder, layer: CaffeLayer): Unit = {
// val donotRegularizeLayers:Boolean = layer.isInstanceOf[BatchNorm] || layer.isInstanceOf[Scale];
val regularizationSource =
if(regularizationType.toLowerCase.equals("l2")) "l2_reg"
else if(regularizationType.toLowerCase.equals("l1")) "l1_reg"
else null
if(regularizationSource == null) {
throw new DMLRuntimeException("Unsupported regularization_type:" + regularizationType + ". Please use either L2 or L1.")
}
if (lambda != 0 && layer.shouldUpdateWeight) {
// Use layer-specific decay multiplier, if param { lr_mult: 1 decay_mult: 1 } is specified in the network file
val hasDecayMult = layer.param.getParamList != null && layer.param.getParamList.size >= 1 && layer.param.getParamList.get(0).hasDecayMult
val newLambda = if(hasDecayMult) layer.param.getParamList.get(0).getDecayMult * lambda else lambda
dmlScript.append("\t").append(layer.dWeight + "_reg = " + regularizationSource + "::backward(" + layer.weight + ", " + newLambda + ")\n")
dmlScript.append("\t").append(layer.dWeight + " = " + layer.dWeight + " + " + layer.dWeight + "_reg\n")
if(layer.shouldUpdateExtraWeight) {
dmlScript.append("\t").append(layer.dExtraWeight + "_reg = " + regularizationSource + "::backward(" + layer.extraWeight + ", " + newLambda + ")\n")
dmlScript.append("\t").append(layer.dExtraWeight + " = " + layer.dExtraWeight + " + " + layer.dExtraWeight + "_reg\n")
}
}
}
}
class LearningRatePolicy(lr_policy: String = "exp", base_lr: Double = 0.01) {
def this(solver: Caffe.SolverParameter) {
this(solver.getLrPolicy, solver.getBaseLr)
if (solver.hasGamma) setGamma(solver.getGamma)
if (solver.hasStepsize) setStepsize(solver.getStepsize)
if (solver.hasPower()) setPower(solver.getPower)
}
var gamma: Double = 0.95
var step: Double = 100000
var power: Double = 0.75
def setGamma(gamma1: Double): Unit = gamma = gamma1
def setStepsize(step1: Double): Unit = step = step1
def setPower(power1: Double): Unit = power = power1
def updateLearningRate(dmlScript: StringBuilder): Unit = {
val new_lr = lr_policy.toLowerCase match {
case "fixed" => base_lr.toString
case "step" => "(" + base_lr + " * " + gamma + " ^ " + " floor(e/" + step + "))"
case "exp" => "(" + base_lr + " * " + gamma + "^e)"
case "inv" => "(" + base_lr + "* (1 + " + gamma + " * e) ^ (-" + power + "))"
case "poly" => "(" + base_lr + " * (1 - e/ max_epochs) ^ " + power + ")"
case "sigmoid" => "(" + base_lr + "( 1/(1 + exp(-" + gamma + "* (e - " + step + "))))"
case _ => throw new DMLRuntimeException("The lr policy is not supported:" + lr_policy)
}
dmlScript.append("lr = " + new_lr + "\n")
}
}
class SGD(regularizationType:String = "L2", lambda: Double = 5e-04, momentum: Double = 0.9) extends CaffeSolver {
/*
* Performs an SGD update with momentum.
*
* In SGD with momentum, we assume that the parameters have a velocity
* that continues with some momentum, and that is influenced by the
* gradient.
*
* Inputs:
* - X: Parameters to update, of shape (any, any).
* - dX: Gradient wrt `X` of a loss function being optimized, of
* same shape as `X`.
* - lr: Learning rate.
* - mu: Momentum value.
* Typical values are in the range of [0.5, 0.99], usually
* started at the lower end and annealed towards the higher end.
* - v: State maintaining the velocity of the parameters `X`, of same
* shape as `X`.
*
* Outputs:
* - X: Updated parameters `X`, of same shape as input `X`.
* - v: Updated velocity of the parameters `X`, of same shape as
* input `X`.
*/
def update(dmlScript: StringBuilder, layer: CaffeLayer): Unit = {
regularization_update(regularizationType, lambda, dmlScript, layer)
if (momentum == 0) {
// Use sgd
if (layer.shouldUpdateWeight) dmlScript.append("\t").append(layer.weight + " = sgd::update(" + commaSep(layer.weight, layer.dWeight, getWeightLr(layer)) + ")\n")
if (layer.shouldUpdateExtraWeight) dmlScript.append("\t").append(layer.extraWeight + " = sgd::update(" + commaSep(layer.extraWeight, layer.dExtraWeight, getWeightLr(layer)) + ")\n")
if (layer.shouldUpdateBias) dmlScript.append("\t").append(layer.bias + " = sgd::update(" + commaSep(layer.bias, layer.dBias, getBiasLr(layer)) + ")\n")
} else {
// Use sgd_momentum
if (layer.shouldUpdateWeight)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.weight, layer.weight + "_v") + "] " +
"= sgd_momentum::update(" + commaSep(layer.weight, layer.dWeight, getWeightLr(layer), momentum.toString, layer.weight + "_v") + ")\n"
)
if (layer.shouldUpdateExtraWeight)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.extraWeight, layer.extraWeight + "_v") + "] " +
"= sgd_momentum::update(" + commaSep(layer.extraWeight, layer.dExtraWeight, getWeightLr(layer), momentum.toString, layer.extraWeight + "_v") + ")\n"
)
if (layer.shouldUpdateBias)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.bias, layer.bias + "_v") + "] " +
"= sgd_momentum::update(" + commaSep(layer.bias, layer.dBias, getBiasLr(layer), momentum.toString, layer.bias + "_v") + ")\n"
)
}
}
def init(dmlScript: StringBuilder, layer: CaffeLayer): Unit =
if (momentum != 0) {
if (layer.shouldUpdateWeight) dmlScript.append(layer.weight + "_v = sgd_momentum::init(" + layer.weight + ")\n")
if (layer.shouldUpdateExtraWeight) dmlScript.append(layer.extraWeight + "_v = sgd_momentum::init(" + layer.extraWeight + ")\n")
if (layer.shouldUpdateBias) dmlScript.append(layer.bias + "_v = sgd_momentum::init(" + layer.bias + ")\n")
}
def sourceFileName: String = if (momentum == 0) "sgd" else "sgd_momentum"
}
class AdaGrad(regularizationType:String = "L2", lambda: Double = 5e-04, epsilon: Double = 1e-6) extends CaffeSolver {
/*
* Performs an Adagrad update.
*
* This is an adaptive learning rate optimizer that maintains the
* sum of squared gradients to automatically adjust the effective
* learning rate.
*
* Reference:
* - Adaptive Subgradient Methods for Online Learning and Stochastic
* Optimization, Duchi et al.
* - http://jmlr.org/papers/v12/duchi11a.html
*
* Inputs:
* - X: Parameters to update, of shape (any, any).
* - dX: Gradient wrt `X` of a loss function being optimized, of
* same shape as `X`.
* - lr: Learning rate.
* - epsilon: Smoothing term to avoid divide by zero errors.
* Typical values are in the range of [1e-8, 1e-4].
* - cache: State that maintains per-parameter sum of squared
* gradients, of same shape as `X`.
*
* Outputs:
* - X: Updated parameters `X`, of same shape as input `X`.
* - cache: State that maintains per-parameter sum of squared
* gradients, of same shape as `X`.
*/
def update(dmlScript: StringBuilder, layer: CaffeLayer): Unit = {
regularization_update(regularizationType, lambda, dmlScript, layer)
if (layer.shouldUpdateWeight)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.weight, layer.weight + "_cache") + "] " +
"= adagrad::update(" + commaSep(layer.weight, layer.dWeight, getWeightLr(layer), epsilon.toString, layer.weight + "_cache") + ")\n"
)
if (layer.shouldUpdateExtraWeight)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.extraWeight, layer.extraWeight + "_cache") + "] " +
"= adagrad::update(" + commaSep(layer.extraWeight, layer.dExtraWeight, getWeightLr(layer), epsilon.toString, layer.extraWeight + "_cache") + ")\n"
)
if (layer.shouldUpdateBias)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.bias, layer.bias + "_cache") + "] " +
"= adagrad::update(" + commaSep(layer.bias, layer.dBias, getBiasLr(layer), epsilon.toString, layer.bias + "_cache") + ")\n"
)
}
def init(dmlScript: StringBuilder, layer: CaffeLayer): Unit = {
if (layer.shouldUpdateWeight) dmlScript.append(layer.weight + "_cache = adagrad::init(" + layer.weight + ")\n")
if (layer.shouldUpdateExtraWeight) dmlScript.append(layer.extraWeight + "_cache = adagrad::init(" + layer.extraWeight + ")\n")
if (layer.shouldUpdateBias) dmlScript.append(layer.bias + "_cache = adagrad::init(" + layer.bias + ")\n")
}
def sourceFileName: String = "adagrad"
}
class Adam(regularizationType:String = "L2", lambda: Double = 5e-04, momentum:Double = 0.9, momentum2:Double = 0.999, delta:Double = 1e-8) extends CaffeSolver {
/*
* Performs an Adam update.
*
* Reference:
* - Adam: A Method for Stochastic Optimization, Kingma, Ba.
* - http://arxiv.org/abs/1412.6980
*
* Inputs:
* - X: Parameters to update, of shape (any, any).
* - dX: Gradient wrt `X` of a loss function being optimized, of
* same shape as `X`.
* - lr: Learning rate. Recommended value is 0.001.
* - beta1: Exponential decay rate for the 1st moment estimates.
* Recommended value is 0.9.
* - beta2: Exponential decay rate for the 2nd moment estimates.
* Recommended value is 0.999.
* - epsilon: Smoothing term to avoid divide by zero errors.
* Recommended value is 1e-8.
* - t: Timestep, starting at 0.
* - m: State containing the 1st moment (mean) estimate by
* maintaining exponential moving averages of the gradients, of
* same shape as `X`.
* - v: State containing the 2nd raw moment (uncentered variance)
* estimate by maintaining exponential moving averages of the
* squared gradients, of same shape as `X`.
*
* Outputs:
* - X: Updated parameters `X`, of same shape as input `X`.
* - m: Updated state containing the 1st moment (mean) estimate by
* maintaining exponential moving averages of the gradients, of
* same shape as `X`.
* - v: Updated state containing the 2nd raw moment (uncentered
* variance) estimate by maintaining exponential moving averages
* of the squared gradients, of same shape as `X`.
*/
def update(dmlScript: StringBuilder, layer: CaffeLayer): Unit = {
regularization_update(regularizationType, lambda, dmlScript, layer)
val t:String = "iter - 1" // since iter starts with 0
// X, dX, double lr, double beta1, double beta2, epsilon, int t, matrix[double] m, matrix[double] v
if (layer.shouldUpdateWeight)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.weight, layer.weight + "_m", layer.weight + "_v") + "] " +
"= adam::update(" + commaSep(layer.weight, layer.dWeight, getWeightLr(layer),
momentum.toString, momentum2.toString, delta.toString, t,
layer.weight + "_m", layer.weight + "_v") + ")\n"
)
if (layer.shouldUpdateExtraWeight)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.extraWeight, layer.extraWeight + "_m", layer.extraWeight + "_v") + "] " +
"= adam::update(" + commaSep(layer.extraWeight, layer.dExtraWeight, getWeightLr(layer),
momentum.toString, momentum2.toString, delta.toString, t,
layer.extraWeight + "_m", layer.extraWeight + "_v") + ")\n"
)
if (layer.shouldUpdateBias)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.bias, layer.bias + "_m", layer.bias + "_v") + "] " +
"= adam::update(" + commaSep(layer.bias, layer.dBias, getBiasLr(layer),
momentum.toString, momentum2.toString, delta.toString, t,
layer.weight + "_m", layer.weight + "_v") + ")\n"
)
}
def init(dmlScript: StringBuilder, layer: CaffeLayer): Unit = {
if (layer.shouldUpdateWeight) dmlScript.append("[ " + layer.weight + "_m, " + layer.weight + "_v ] = adam::init(" + layer.weight + ")\n")
if (layer.shouldUpdateExtraWeight) dmlScript.append("[ " + layer.extraWeight + "_m, " + layer.extraWeight + "_v ] = adam::init(" + layer.extraWeight + ")\n")
if (layer.shouldUpdateBias) dmlScript.append("[ " + layer.bias + "_m, " + layer.bias + "_v ] = adam::init(" + layer.bias + ")\n")
}
def sourceFileName: String = "adam"
}
class Nesterov(regularizationType:String = "L2", lambda: Double = 5e-04, momentum: Double = 0.9) extends CaffeSolver {
/*
* Performs an SGD update with Nesterov momentum.
*
* As with regular SGD with momentum, in SGD with Nesterov momentum,
* we assume that the parameters have a velocity that continues
* with some momentum, and that is influenced by the gradient.
* In this view specifically, we perform the position update from the
* position that the momentum is about to carry the parameters to,
* rather than from the previous position. Additionally, we always
* store the parameters in their position after momentum.
*
* Reference:
* - Advances in optimizing Recurrent Networks, Bengio et al.,
* section 3.5.
* - http://arxiv.org/abs/1212.0901
*
* Inputs:
* - X: Parameters to update, of shape (any, any).
* - dX: Gradient wrt `X` of a loss function being optimized, of
* same shape as `X`.
* - lr: Learning rate.
* - mu: Momentum value.
* Typical values are in the range of [0.5, 0.99], usually
* started at the lower end and annealed towards the higher end.
* - v: State maintaining the velocity of the parameters `X`, of same
* shape as `X`.
*
* Outputs:
* - X: Updated parameters X, of same shape as input X.
* - v: Updated velocity of the parameters X, of same shape as
* input v.
*/
def update(dmlScript: StringBuilder, layer: CaffeLayer): Unit = {
val fn = if (Caffe2DML.USE_NESTEROV_UDF) "update_nesterov" else "sgd_nesterov::update"
val lastParameter = if (Caffe2DML.USE_NESTEROV_UDF) (", " + lambda) else ""
if (!Caffe2DML.USE_NESTEROV_UDF) {
regularization_update(regularizationType, lambda, dmlScript, layer)
}
if (layer.shouldUpdateWeight)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.weight, layer.weight + "_v") + "] " +
"= " + fn + "(" + commaSep(layer.weight, layer.dWeight, getWeightLr(layer), momentum.toString, layer.weight + "_v") + lastParameter + ")\n"
)
if (layer.shouldUpdateExtraWeight)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.extraWeight, layer.extraWeight + "_v") + "] " +
"= " + fn + "(" + commaSep(layer.extraWeight, layer.dExtraWeight, getWeightLr(layer), momentum.toString, layer.extraWeight + "_v") + lastParameter + ")\n"
)
if (layer.shouldUpdateBias)
dmlScript
.append("\t")
.append(
"[" + commaSep(layer.bias, layer.bias + "_v") + "] " +
"= " + fn + "(" + commaSep(layer.bias, layer.dBias, getBiasLr(layer), momentum.toString, layer.bias + "_v") + lastParameter + ")\n"
)
}
def init(dmlScript: StringBuilder, layer: CaffeLayer): Unit = {
if (layer.shouldUpdateWeight) dmlScript.append(layer.weight + "_v = sgd_nesterov::init(" + layer.weight + ")\n")
if (layer.shouldUpdateExtraWeight) dmlScript.append(layer.extraWeight + "_v = sgd_nesterov::init(" + layer.extraWeight + ")\n")
if (layer.shouldUpdateBias) dmlScript.append(layer.bias + "_v = sgd_nesterov::init(" + layer.bias + ")\n")
}
def sourceFileName: String = "sgd_nesterov"
}