@Test
public void testAdamUpdater() {
INDArray m, v;
double lr = 0.01;
int iteration = 0;
double beta1 = 0.8;
double beta2 = 0.888;
NeuralNetConfiguration conf =
new NeuralNetConfiguration.Builder()
.learningRate(lr)
.iterations(iteration)
.adamMeanDecay(beta1)
.adamVarDecay(beta2)
.layer(
new DenseLayer.Builder()
.nIn(nIn)
.nOut(nOut)
.updater(org.deeplearning4j.nn.conf.Updater.ADAM)
.build())
.build();
int numParams = LayerFactories.getFactory(conf).initializer().numParams(conf, true);
INDArray params = Nd4j.create(1, numParams);
Layer layer = LayerFactories.getFactory(conf).create(conf, null, 0, params, true);
Updater updater = UpdaterCreator.getUpdater(layer);
int updaterStateSize = updater.stateSizeForLayer(layer);
INDArray updaterState = Nd4j.create(1, updaterStateSize);
updater.setStateViewArray(layer, updaterState, true);
updater.update(layer, gradient, iteration, 1);
double beta1t = FastMath.pow(beta1, iteration);
double beta2t = FastMath.pow(beta2, iteration);
double alphat = lr * FastMath.sqrt(1 - beta2t) / (1 - beta1t);
if (Double.isNaN(alphat) || alphat == 0.0) alphat = epsilon;
Gradient gradientDup = new DefaultGradient();
gradientDup.setGradientFor(DefaultParamInitializer.WEIGHT_KEY, weightGradient);
gradientDup.setGradientFor(DefaultParamInitializer.BIAS_KEY, biasGradient);
for (Map.Entry<String, INDArray> entry : gradientDup.gradientForVariable().entrySet()) {
val = entry.getValue();
m = Nd4j.zeros(val.shape());
v = Nd4j.zeros(val.shape());
m.muli(beta1).addi(val.mul(1.0 - beta1));
v.muli(beta2).addi(val.mul(val).mul(1.0 - beta2));
gradExpected = m.mul(alphat).divi(Transforms.sqrt(v).addi(epsilon));
if (!gradExpected.equals(gradient.getGradientFor(entry.getKey()))) {
System.out.println(Arrays.toString(gradExpected.dup().data().asFloat()));
System.out.println(
Arrays.toString(gradient.getGradientFor(entry.getKey()).dup().data().asFloat()));
}
assertEquals(gradExpected, gradient.getGradientFor(entry.getKey()));
}
assertEquals(beta1, layer.conf().getLayer().getAdamMeanDecay(), 1e-4);
assertEquals(beta2, layer.conf().getLayer().getAdamVarDecay(), 1e-4);
}
@Test
public void testRMSPropUpdater() {
double lr = 0.01;
double rmsDecay = 0.25;
Map<String, INDArray> lastG = new HashMap<>();
NeuralNetConfiguration conf =
new NeuralNetConfiguration.Builder()
.learningRate(lr)
.rmsDecay(rmsDecay)
.layer(
new DenseLayer.Builder()
.nIn(nIn)
.nOut(nOut)
.updater(org.deeplearning4j.nn.conf.Updater.RMSPROP)
.build())
.build();
int numParams = LayerFactories.getFactory(conf).initializer().numParams(conf, true);
INDArray params = Nd4j.create(1, numParams);
Layer layer = LayerFactories.getFactory(conf).create(conf, null, 0, params, true);
Updater updater = UpdaterCreator.getUpdater(layer);
int updaterStateSize = updater.stateSizeForLayer(layer);
INDArray updaterState = Nd4j.create(1, updaterStateSize);
updater.setStateViewArray(layer, updaterState, true);
updater.update(layer, gradient, -1, 1);
Gradient gradientDup = new DefaultGradient();
gradientDup.setGradientFor(DefaultParamInitializer.WEIGHT_KEY, weightGradient.dup());
gradientDup.setGradientFor(DefaultParamInitializer.BIAS_KEY, biasGradient.dup());
for (Map.Entry<String, INDArray> entry : gradientDup.gradientForVariable().entrySet()) {
key = entry.getKey();
val = entry.getValue();
INDArray lastGTmp = lastG.get(key);
if (lastGTmp == null) lastGTmp = Nd4j.zeros(val.shape());
lastGTmp.muli(rmsDecay).addi(val.mul(val).muli(1 - rmsDecay));
gradExpected = val.mul(lr).div(Transforms.sqrt(lastGTmp.add(Nd4j.EPS_THRESHOLD)));
assertEquals(gradExpected, gradient.getGradientFor(entry.getKey()));
lastG.put(key, lastGTmp);
}
assertEquals(rmsDecay, layer.conf().getLayer().getRmsDecay(), 1e-4);
}
@Test
public void testNestorovsUpdater() {
double lr = 1e-2;
double mu = 0.6;
INDArray v, vPrev;
NeuralNetConfiguration conf =
new NeuralNetConfiguration.Builder()
.learningRate(lr)
.momentum(mu)
.layer(
new DenseLayer.Builder()
.nIn(nIn)
.nOut(nOut)
.updater(org.deeplearning4j.nn.conf.Updater.NESTEROVS)
.build())
.build();
int numParams = LayerFactories.getFactory(conf).initializer().numParams(conf, true);
INDArray params = Nd4j.create(1, numParams);
Layer layer = LayerFactories.getFactory(conf).create(conf, null, 0, params, true);
Updater updater = UpdaterCreator.getUpdater(layer);
int updaterStateSize = updater.stateSizeForLayer(layer);
INDArray updaterState = Nd4j.create(1, updaterStateSize);
updater.setStateViewArray(layer, updaterState, true);
updater.update(layer, gradient, -1, 1);
Gradient gradientDup = new DefaultGradient();
gradientDup.setGradientFor(DefaultParamInitializer.WEIGHT_KEY, weightGradient.dup());
gradientDup.setGradientFor(DefaultParamInitializer.BIAS_KEY, biasGradient.dup());
for (Map.Entry<String, INDArray> entry : gradientDup.gradientForVariable().entrySet()) {
val = entry.getValue();
v = Nd4j.zeros(val.shape());
vPrev = v;
v = vPrev.mul(mu).subi(val.mul(lr));
gradExpected = vPrev.muli(mu).addi(v.mul(-mu - 1));
assertEquals(gradExpected, gradient.getGradientFor(entry.getKey()));
}
assertEquals(mu, layer.conf().getLayer().getMomentum(), 1e-4);
}
@Test
public void testAdaGradUpdater() {
double lr = 1e-2;
NeuralNetConfiguration conf =
new NeuralNetConfiguration.Builder()
.learningRate(lr)
.layer(
new DenseLayer.Builder()
.nIn(nIn)
.nOut(nOut)
.updater(org.deeplearning4j.nn.conf.Updater.ADAGRAD)
.build())
.build();
int numParams = LayerFactories.getFactory(conf).initializer().numParams(conf, true);
INDArray params = Nd4j.create(1, numParams);
Layer layer = LayerFactories.getFactory(conf).create(conf, null, 0, params, true);
Updater updater = UpdaterCreator.getUpdater(layer);
int updaterStateSize = updater.stateSizeForLayer(layer);
INDArray updaterState = Nd4j.create(1, updaterStateSize);
updater.setStateViewArray(layer, updaterState, true);
updater.update(layer, gradient, -1, 1);
Gradient gradientDup = new DefaultGradient();
gradientDup.setGradientFor(DefaultParamInitializer.WEIGHT_KEY, weightGradient);
gradientDup.setGradientFor(DefaultParamInitializer.BIAS_KEY, biasGradient);
for (Map.Entry<String, INDArray> entry : gradientDup.gradientForVariable().entrySet()) {
val = entry.getValue();
gradExpected = Transforms.sqrt(val.mul(val).add(epsilon)).rdiv(lr).mul(val);
assertEquals(gradExpected, gradient.getGradientFor(entry.getKey()));
}
assertEquals(lr, layer.conf().getLayer().getLearningRate(), 1e-4);
}
@Test
public void testAdaDeltaUpdate() {
INDArray dxSquared;
Map<String, INDArray> msg = new HashMap<>();
Map<String, INDArray> msdx = new HashMap<>();
double rho = 0.85;
NeuralNetConfiguration conf =
new NeuralNetConfiguration.Builder()
.rho(rho)
.layer(
new DenseLayer.Builder()
.nIn(nIn)
.nOut(nOut)
.updater(org.deeplearning4j.nn.conf.Updater.ADADELTA)
.build())
.build();
int numParams = LayerFactories.getFactory(conf).initializer().numParams(conf, true);
INDArray params = Nd4j.create(1, numParams);
Layer layer = LayerFactories.getFactory(conf).create(conf, null, 0, params, true);
Updater updater = UpdaterCreator.getUpdater(layer);
int updaterStateSize = updater.stateSizeForLayer(layer);
INDArray updaterState = Nd4j.create(1, updaterStateSize);
updater.setStateViewArray(layer, updaterState, true);
Gradient gradientDup = new DefaultGradient();
gradientDup.setGradientFor(DefaultParamInitializer.WEIGHT_KEY, weightGradient.dup());
gradientDup.setGradientFor(DefaultParamInitializer.BIAS_KEY, biasGradient.dup());
for (int i = 0; i < 2; i++) {
updater.update(layer, gradient, i, 1);
// calculations for one iteration / update
for (Map.Entry<String, INDArray> entry : gradientDup.gradientForVariable().entrySet()) {
key = entry.getKey();
val = entry.getValue();
INDArray msgTmp = msg.get(key);
INDArray msdxTmp = msdx.get(key);
if (msgTmp == null) {
msgTmp = Nd4j.zeros(val.shape());
msdxTmp = Nd4j.zeros(val.shape());
}
msgTmp.muli(rho);
msgTmp.addi(1 - rho).muli(val.mul(val));
gradExpected =
Transforms.sqrt(msdxTmp.add(Nd4j.EPS_THRESHOLD))
.divi(Transforms.sqrt(msgTmp.add(Nd4j.EPS_THRESHOLD)))
.muli(val);
gradientDup.setGradientFor(key, gradExpected);
assertEquals(gradExpected, gradient.getGradientFor(entry.getKey()));
msdxTmp.muli(rho);
dxSquared = gradExpected.mul(gradExpected);
msdxTmp.addi(dxSquared.muli(1 - rho));
msg.put(key, msgTmp);
msdx.put(key, msdxTmp);
}
assertEquals(rho, layer.conf().getLayer().getRho(), 1e-4);
}
}