com.simiacryptus.mindseye.network.util.SigmoidTreeNetwork Maven / Gradle / Ivy
/*
* Copyright (c) 2019 by Andrew Charneski.
*
* The author 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 com.simiacryptus.mindseye.network.util;
import com.google.gson.JsonObject;
import com.simiacryptus.mindseye.lang.DataSerializer;
import com.simiacryptus.mindseye.lang.Layer;
import com.simiacryptus.mindseye.layers.java.*;
import com.simiacryptus.mindseye.network.DAGNetwork;
import com.simiacryptus.mindseye.network.DAGNode;
import com.simiacryptus.util.FastRandom;
import javax.annotation.Nonnull;
import javax.annotation.Nullable;
import java.util.List;
import java.util.Map;
import java.util.UUID;
@SuppressWarnings("serial")
public class SigmoidTreeNetwork extends DAGNetwork implements EvolvingNetwork {
private final boolean multigate = false;
double initialFuzzyCoeff = 1e-8;
@Nullable
private Layer alpha = null;
@Nullable
private Layer alphaBias = null;
@Nullable
private Layer beta = null;
@Nullable
private Layer betaBias = null;
@Nullable
private Layer gate = null;
@Nullable
private Layer gateBias = null;
@Nullable
private DAGNode head = null;
@Nullable
private NodeMode mode = null;
private boolean skipChildStage = true;
private boolean skipFuzzy = false;
protected SigmoidTreeNetwork(@Nonnull final JsonObject json, Map rs) {
super(json, rs);
head = getNodeById(UUID.fromString(json.get("head").getAsString()));
Map layersById = getLayersById();
if (json.get("alpha") != null) {
alpha = layersById.get(UUID.fromString(json.get("alpha").getAsString()));
}
if (json.get("alphaBias") != null) {
alphaBias = layersById.get(UUID.fromString(json.get("alphaBias").getAsString()));
}
if (json.get("beta") != null) {
beta = layersById.get(UUID.fromString(json.get("beta").getAsString()));
}
if (json.get("betaBias") != null) {
betaBias = layersById.get(UUID.fromString(json.get("betaBias").getAsString()));
}
if (json.get("gate") != null) {
gate = layersById.get(UUID.fromString(json.get("gate").getAsString()));
}
if (json.get("gateBias") != null) {
gate = layersById.get(UUID.fromString(json.get("gateBias").getAsString()));
}
setSkipChildStage(json.get("skipChildStage") != null ? json.get("skipChildStage").getAsBoolean() : skipChildStage());
setSkipFuzzy(json.get("skipFuzzy") != null ? json.get("skipFuzzy").getAsBoolean() : isSkipFuzzy());
mode = NodeMode.valueOf(json.get("mode").getAsString());
}
public SigmoidTreeNetwork(final Layer alpha, final Layer alphaBias) {
super(1);
this.alpha = alpha;
this.alphaBias = alphaBias;
mode = NodeMode.Linear;
}
public static SigmoidTreeNetwork fromJson(@Nonnull final JsonObject json, Map rs) {
return new SigmoidTreeNetwork(json, rs);
}
@Override
protected void _free() {
head.freeRef();
alpha.freeRef();
alphaBias.freeRef();
beta.freeRef();
betaBias.freeRef();
gate.freeRef();
super._free();
}
public void copyState(@Nonnull final Layer from, @Nonnull final Layer to) {
@Nullable final List alphaState = from.state();
@Nullable final List betaState = to.state();
for (int i = 0; i < alphaState.size(); i++) {
final double[] betaBuffer = betaState.get(i);
final double[] alphaBuffer = alphaState.get(i);
System.arraycopy(alphaBuffer, 0, betaBuffer, 0, alphaBuffer.length);
}
}
@Nullable
@Override
public synchronized DAGNode getHead() {
if (null == head) {
synchronized (this) {
if (null == head) {
reset();
final DAGNode input = getInput(0);
switch (getMode()) {
case Linear:
head = add(alpha.setFrozen(false), add(alphaBias.setFrozen(false), input));
break;
case Fuzzy: {
final DAGNode gateNode = add(gate.setFrozen(false), null != gateBias ? add(gateBias.setFrozen(false), input) : input);
head = add(new ProductInputsLayer(),
add(alpha.setFrozen(false), add(alphaBias.setFrozen(false), input)),
add(new LinearActivationLayer().setScale(2).freeze(),
add(new SigmoidActivationLayer().setBalanced(false), gateNode))
);
break;
}
case Bilinear: {
final DAGNode gateNode = add(gate.setFrozen(false), null != gateBias ? add(gateBias.setFrozen(false), input) : input);
head = add(new SumInputsLayer(),
add(new ProductInputsLayer(),
add(alpha.setFrozen(false), add(alphaBias.setFrozen(false), input)),
add(new SigmoidActivationLayer().setBalanced(false), gateNode)
),
add(new ProductInputsLayer(),
add(beta.setFrozen(false), add(betaBias.setFrozen(false), input)),
add(new SigmoidActivationLayer().setBalanced(false),
add(new LinearActivationLayer().setScale(-1).freeze(), gateNode))
));
break;
}
case Final:
final DAGNode gateNode = add(gate.setFrozen(false), null != gateBias ? add(gateBias.setFrozen(false), input) : input);
head = add(new SumInputsLayer(),
add(new ProductInputsLayer(),
add(alpha, input),
add(new SigmoidActivationLayer().setBalanced(false), gateNode)
),
add(new ProductInputsLayer(),
add(beta, input),
add(new SigmoidActivationLayer().setBalanced(false),
add(new LinearActivationLayer().setScale(-1).freeze(), gateNode))
));
break;
}
}
}
}
head.addRef();
return head;
}
@Override
public JsonObject getJson(Map resources, DataSerializer dataSerializer) {
assertConsistent();
UUID headId = getHeadId();
final JsonObject json = super.getJson(resources, dataSerializer);
json.addProperty("head", headId.toString());
if (null != alpha) {
json.addProperty("alpha", alpha.getId().toString());
}
if (null != alphaBias) {
json.addProperty("alphaBias", alpha.getId().toString());
}
if (null != beta) {
json.addProperty("beta", beta.getId().toString());
}
if (null != betaBias) {
json.addProperty("betaBias", beta.getId().toString());
}
if (null != gate) {
json.addProperty("gate", gate.getId().toString());
}
if (null != gateBias) {
json.addProperty("gateBias", gate.getId().toString());
}
json.addProperty("mode", getMode().name());
json.addProperty("skipChildStage", skipChildStage());
json.addProperty("skipFuzzy", isSkipFuzzy());
assert null != Layer.fromJson(json) : "Smoke apply deserialization";
return json;
}
@Nullable
public NodeMode getMode() {
return mode;
}
public boolean isSkipFuzzy() {
return skipFuzzy;
}
@Nonnull
public SigmoidTreeNetwork setSkipFuzzy(final boolean skipFuzzy) {
this.skipFuzzy = skipFuzzy;
return this;
}
@Override
public void nextPhase() {
switch (getMode()) {
case Linear: {
head = null;
@Nonnull final FullyConnectedLayer alpha = (FullyConnectedLayer) this.alpha;
//alphaList.weights.scale(2);
gate = new FullyConnectedLayer(alpha.inputDims, multigate ? alpha.outputDims : new int[]{1});
gateBias = new BiasLayer(alpha.inputDims);
mode = NodeMode.Fuzzy;
break;
}
case Fuzzy: {
head = null;
@Nullable final FullyConnectedLayer alpha = (FullyConnectedLayer) this.alpha;
@Nonnull final BiasLayer alphaBias = (BiasLayer) this.alphaBias;
beta = new FullyConnectedLayer(alpha.inputDims, alpha.outputDims).set(() -> {
return initialFuzzyCoeff * (FastRandom.INSTANCE.random() - 0.5);
});
betaBias = new BiasLayer(alphaBias.bias.length());
copyState(alpha, beta);
copyState(alphaBias, betaBias);
mode = NodeMode.Bilinear;
if (isSkipFuzzy()) {
nextPhase();
}
break;
}
case Bilinear:
head = null;
alpha = new SigmoidTreeNetwork(alpha, alphaBias);
if (skipChildStage()) {
((SigmoidTreeNetwork) alpha).nextPhase();
}
beta = new SigmoidTreeNetwork(beta, betaBias);
if (skipChildStage()) {
((SigmoidTreeNetwork) beta).nextPhase();
}
mode = NodeMode.Final;
break;
case Final:
@Nonnull final SigmoidTreeNetwork alpha = (SigmoidTreeNetwork) this.alpha;
@Nonnull final SigmoidTreeNetwork beta = (SigmoidTreeNetwork) this.beta;
alpha.nextPhase();
beta.nextPhase();
break;
}
}
@Nonnull
public SigmoidTreeNetwork setSkipChildStage(final boolean skipChildStage) {
this.skipChildStage = skipChildStage;
return this;
}
public boolean skipChildStage() {
return skipChildStage;
}
public enum NodeMode {
Bilinear,
Final,
Fuzzy,
Linear
}
}
© 2015 - 2025 Weber Informatics LLC | Privacy Policy