// note precision is off on this test but the numbers are close
// investigation in a future release should determine how to resolve
@Test
public void testCalculateDeltaContained() {
Layer layer = getContainedConfig();
INDArray input = getContainedData();
INDArray col = getContainedCol();
INDArray epsilon = Nd4j.ones(1, 2, 4, 4);
INDArray expectedOutput =
Nd4j.create(
new double[] {
0.02036651, 0.02036651, 0.02036651, 0.02036651, 0.00039383,
0.00039383, 0.00039383, 0.00039383, 0.02036651, 0.02036651,
0.02036651, 0.02036651, 0.00039383, 0.00039383, 0.00039383,
0.00039383, 0.02036651, 0.02036651, 0.02036651, 0.02036651,
0.00039383, 0.00039383, 0.00039383, 0.00039383, 0.02036651,
0.02036651, 0.02036651, 0.02036651, 0.00039383, 0.00039383,
0.00039383, 0.00039383
},
new int[] {1, 2, 4, 4});
layer.setInput(input);
org.deeplearning4j.nn.layers.convolution.ConvolutionLayer layer2 =
(org.deeplearning4j.nn.layers.convolution.ConvolutionLayer) layer;
layer2.setCol(col);
INDArray delta = layer2.calculateDelta(epsilon);
assertArrayEquals(expectedOutput.shape(), delta.shape());
assertEquals(expectedOutput, delta);
}
// TODO remove/move, technically this is testing Nd4j functionality
@Test
public void testCreateFeatureMapMethod() {
Layer layer = getContainedConfig();
INDArray input = getContainedData();
int inputWidth = input.shape()[0];
int featureMapWidth =
(inputWidth + layer.conf().getPadding()[0] * 2 - layer.conf().getKernelSize()[0])
/ layer.conf().getStride()[0]
+ 1;
INDArray expectedOutput =
Nd4j.create(
new double[] {
1, 1, 1, 1, 3, 3, 3, 3, 1, 1, 1, 1, 3, 3, 3, 3, 1, 1, 1, 1, 3, 3, 3,
3, 1, 1, 1, 1, 3, 3, 3, 3, 2, 2, 2, 2, 4, 4, 4, 4, 2, 2, 2, 2, 4, 4,
4, 4, 2, 2, 2, 2, 4, 4, 4, 4, 2, 2, 2, 2, 4, 4, 4, 4
},
new int[] {1, 1, 2, 2, 4, 4});
layer.setInput(input);
org.deeplearning4j.nn.layers.convolution.ConvolutionLayer layer2 =
(org.deeplearning4j.nn.layers.convolution.ConvolutionLayer) layer;
INDArray featureMaps = layer2.createFeatureMapColumn();
assertEquals(featureMapWidth, featureMaps.shape()[4]);
assertEquals(expectedOutput.shape(), featureMaps.shape());
assertEquals(expectedOutput, featureMaps);
}
// note precision is off on this test but the numbers are close
// investigation in a future release should determine how to resolve
@Test
public void testBackpropResultsContained() {
Layer layer = getContainedConfig();
INDArray input = getContainedData();
INDArray col = getContainedCol();
INDArray epsilon = Nd4j.ones(1, 2, 4, 4);
INDArray expectedBiasGradient =
Nd4j.create(new double[] {0.16608272, 0.16608272}, new int[] {1, 2});
INDArray expectedWeightGradient =
Nd4j.create(
new double[] {
0.17238397,
0.17238397,
0.33846668,
0.33846668,
0.17238397,
0.17238397,
0.33846668,
0.33846668
},
new int[] {2, 1, 2, 2});
INDArray expectedEpsilon =
Nd4j.create(
new double[] {
0.00039383, 0.00039383, 0.00039383, 0.00039383, 0.00039383,
0.00039383, 0., 0., 0.00039383, 0.00039383,
0.00039383, 0.00039383, 0.00039383, 0.00039383, 0.,
0., 0.02036651, 0.02036651, 0.02036651, 0.02036651,
0.02036651, 0.02036651, 0., 0., 0.02036651,
0.02036651, 0.02036651, 0.02036651, 0.02036651, 0.02036651,
0., 0., 0.00039383, 0.00039383, 0.00039383,
0.00039383, 0.00039383, 0.00039383, 0., 0.,
0.00039383, 0.00039383, 0.00039383, 0.00039383, 0.00039383,
0.00039383, 0., 0., 0., 0.,
0., 0., 0., 0., 0.,
0., 0., 0., 0., 0.,
0., 0., 0., 0.
},
new int[] {1, 1, 8, 8});
layer.setInput(input);
org.deeplearning4j.nn.layers.convolution.ConvolutionLayer layer2 =
(org.deeplearning4j.nn.layers.convolution.ConvolutionLayer) layer;
layer2.setCol(col);
Pair<Gradient, INDArray> pair = layer2.backpropGradient(epsilon);
assertArrayEquals(expectedEpsilon.shape(), pair.getSecond().shape());
assertArrayEquals(expectedWeightGradient.shape(), pair.getFirst().getGradientFor("W").shape());
assertArrayEquals(expectedBiasGradient.shape(), pair.getFirst().getGradientFor("b").shape());
assertEquals(expectedEpsilon, pair.getSecond());
assertEquals(expectedWeightGradient, pair.getFirst().getGradientFor("W"));
assertEquals(expectedBiasGradient, pair.getFirst().getGradientFor("b"));
}
