[SYSTEMML-1076] [SYSTEMML-1077] Handle empty blocks in maxpool and maxpool_backward
diff --git a/src/main/java/org/apache/sysml/hops/ConvolutionOp.java b/src/main/java/org/apache/sysml/hops/ConvolutionOp.java
index c010de6..dae2f82 100644
--- a/src/main/java/org/apache/sysml/hops/ConvolutionOp.java
+++ b/src/main/java/org/apache/sysml/hops/ConvolutionOp.java
@@ -242,7 +242,7 @@
case MAX_POOLING:
{
ret = new long[3];
- ret[0] = params.N;
+ ret[0] = getInput().get(0)._dim1;
ret[1] = getExtractedVal(params.C, params.P, params.Q);
ret[2] = -1;
break;
@@ -250,32 +250,32 @@
case MAX_POOLING_BACKWARD:
{
ret = new long[3];
- ret[0] = params.N;
- ret[1] = getExtractedVal(params.C, params.H, params.W);
+ ret[0] = getInput().get(0)._dim1;
+ ret[1] = getInput().get(0)._dim2;
ret[2] = -1;
break;
}
case DIRECT_CONV2D:
{
ret = new long[3];
- ret[0] = params.N;
- ret[1] = getExtractedVal(params.K, params.P, params.Q);
+ ret[0] = getInput().get(0)._dim1;
+ ret[1] = getExtractedVal(getInput().get(1)._dim1, params.P, params.Q);
ret[2] = -1;
break;
}
case DIRECT_CONV2D_BACKWARD_FILTER:
{
ret = new long[3];
- ret[0] = params.K;
- ret[1] = getExtractedVal(params.C, params.R, params.S);
+ ret[0] = getInput().get(1)._dim1;
+ ret[1] = getInput().get(1)._dim2;
ret[2] = -1;
break;
}
case DIRECT_CONV2D_BACKWARD_DATA:
{
ret = new long[3];
- ret[0] = params.N;
- ret[1] = getExtractedVal(params.C, params.H, params.W);
+ ret[0] = getInput().get(0)._dim1;
+ ret[1] = getInput().get(0)._dim2;
ret[2] = -1;
break;
}
@@ -408,36 +408,36 @@
{
case MAX_POOLING:
{
- _dim1 = params.N;
+ _dim1 = getInput().get(0)._dim1;
_dim2 = getExtractedVal(params.C, params.P, params.Q);
_nnz = -1; // cannot infer stats
break;
}
case MAX_POOLING_BACKWARD:
{
- _dim1 = params.N;
- _dim2 = getExtractedVal(params.C, params.H, params.W);
+ _dim1 = getInput().get(0)._dim1;
+ _dim2 = getInput().get(0)._dim2;
_nnz = -1;
break;
}
case DIRECT_CONV2D:
{
- _dim1 = params.N;
- _dim2 = getExtractedVal(params.K, params.P, params.Q);
+ _dim1 = getInput().get(0)._dim1;
+ _dim2 = getExtractedVal(getInput().get(1)._dim1, params.P, params.Q);
_nnz = -1; // cannot infer stats
break;
}
case DIRECT_CONV2D_BACKWARD_DATA:
{
- _dim1 = params.N;
- _dim2 = getExtractedVal(params.C, params.H, params.W);
+ _dim1 = getInput().get(0)._dim1;
+ _dim2 = getInput().get(0)._dim2;
_nnz = -1; // cannot infer stats
break;
}
case DIRECT_CONV2D_BACKWARD_FILTER:
{
- _dim1 = params.K;
- _dim2 = getExtractedVal(params.C, params.R, params.S);
+ _dim1 = getInput().get(1)._dim1;
+ _dim2 = getInput().get(1)._dim2;
_nnz = -1; // cannot infer stats
break;
}
diff --git a/src/main/java/org/apache/sysml/parser/BuiltinFunctionExpression.java b/src/main/java/org/apache/sysml/parser/BuiltinFunctionExpression.java
index bf31347..0ab2511 100644
--- a/src/main/java/org/apache/sysml/parser/BuiltinFunctionExpression.java
+++ b/src/main/java/org/apache/sysml/parser/BuiltinFunctionExpression.java
@@ -24,6 +24,7 @@
import java.util.HashSet;
import org.apache.sysml.parser.LanguageException.LanguageErrorCodes;
+import org.apache.sysml.runtime.util.ConvolutionUtils;
public class BuiltinFunctionExpression extends DataIdentifier
{
@@ -1121,14 +1122,57 @@
// conv2d_backward_filter and conv2d_backward_data
Expression input = _args[0]; // For conv2d_backward_filter, this is input and for conv2d_backward_data, this is filter
+ Expression filter = null;
if(!(this.getOpCode() == BuiltinFunctionOp.MAX_POOL || this.getOpCode() == BuiltinFunctionOp.AVG_POOL)) {
- Expression filter = _args[1]; // For conv2d_backward functions, this is dout
+ filter = _args[1]; // For conv2d_backward functions, this is dout
checkMatrixParam(filter);
}
output.setDataType(DataType.MATRIX);
output.setValueType(ValueType.DOUBLE);
output.setBlockDimensions(input.getOutput().getRowsInBlock(), input.getOutput().getColumnsInBlock());
-
+ // stride1, stride2, padding1, padding2, numImg, numChannels, imgSize, imgSize,
+ // filter_shape1=1, filter_shape2=1, filterSize/poolSize1, filterSize/poolSize1
+ if(this.getOpCode() == BuiltinFunctionOp.MAX_POOL_BACKWARD ||
+ this.getOpCode() == BuiltinFunctionOp.CONV2D_BACKWARD_DATA) {
+ output.setDimensions(input.getOutput().getDim1(), input.getOutput().getDim2());
+ }
+ else if(this.getOpCode() == BuiltinFunctionOp.CONV2D_BACKWARD_FILTER) {
+ output.setDimensions(filter.getOutput().getDim1(), filter.getOutput().getDim2());
+ }
+ else if(this.getOpCode() == BuiltinFunctionOp.CONV2D || this.getOpCode() == BuiltinFunctionOp.MAX_POOL) {
+ try {
+ int start = 1;
+ if(this.getOpCode() == BuiltinFunctionOp.CONV2D) {
+ start = 2;
+ }
+ long stride_h = (long) getDoubleValue(_args[start++]);
+ long stride_w = (long) getDoubleValue(_args[start++]);
+ long pad_h = (long) getDoubleValue(_args[start++]);
+ long pad_w = (long) getDoubleValue(_args[start++]);
+ start++;
+ long C = (long) getDoubleValue(_args[start++]);
+ long H = (long) getDoubleValue(_args[start++]);
+ long W = (long) getDoubleValue(_args[start++]);
+ long K = -1;
+ if(this.getOpCode() == BuiltinFunctionOp.CONV2D) {
+ K = (long) getDoubleValue(_args[start]);
+ }
+ start++; start++;
+ long R = (long) getDoubleValue(_args[start++]);
+ long S = (long) getDoubleValue(_args[start++]);
+ long P = ConvolutionUtils.getP(H, R, stride_h, pad_h);
+ long Q = ConvolutionUtils.getP(W, S, stride_w, pad_w);
+ if(this.getOpCode() == BuiltinFunctionOp.CONV2D)
+ output.setDimensions(input.getOutput().getDim1(), K*P*Q);
+ else
+ output.setDimensions(input.getOutput().getDim1(), C*P*Q);
+ }
+ catch(Exception e) {
+ output.setDimensions(input.getOutput().getDim1(), -1); // To make sure that output dimensions are not incorrect
+ }
+ }
+ else
+ throw new LanguageException("Unsupported op: " + this.getOpCode());
checkMatrixParam(input);
break;
}
diff --git a/src/main/java/org/apache/sysml/runtime/instructions/cp/ConvolutionCPInstruction.java b/src/main/java/org/apache/sysml/runtime/instructions/cp/ConvolutionCPInstruction.java
index 3885720..91c74f0 100644
--- a/src/main/java/org/apache/sysml/runtime/instructions/cp/ConvolutionCPInstruction.java
+++ b/src/main/java/org/apache/sysml/runtime/instructions/cp/ConvolutionCPInstruction.java
@@ -188,8 +188,8 @@
// without somewhat expensive HashMap checks
outputBlock = getDenseOutputBlock(ec, N, C*P*Q, true);
params.setReuseNonZeroedOutput(_reuseNonZeroedOutput);
+ LibMatrixDNN.maxpooling(matBlock, outputBlock, params);
}
- LibMatrixDNN.maxpooling(matBlock, outputBlock, params);
}
else if (instOpcode.equalsIgnoreCase("maxpooling_backward")) {
MatrixBlock dout = ec.getMatrixInput(_in2.getName());
@@ -201,8 +201,8 @@
// without somewhat expensive HashMap checks
outputBlock = getDenseOutputBlock(ec, N, C*H*W, false);
params.setReuseNonZeroedOutput(_reuseNonZeroedOutput);
+ LibMatrixDNN.maxpooling_backward(matBlock, dout, outputBlock, params);
}
- LibMatrixDNN.maxpooling_backward(matBlock, dout, outputBlock, params);
ec.releaseMatrixInput(_in2.getName());
}
else if (instOpcode.equalsIgnoreCase("conv2d")) {
diff --git a/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixDNN.java b/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixDNN.java
index e785112..cbb846c 100644
--- a/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixDNN.java
+++ b/src/main/java/org/apache/sysml/runtime/matrix/data/LibMatrixDNN.java
@@ -735,13 +735,15 @@
outputBlock.setNonZeros(params.outputNNZ.get());
}
- private static void doPooling(int n, int c, ConvolutionParameters params) {
+ private static void doPooling(int n, int c, ConvolutionParameters params) throws DMLRuntimeException {
double [] inputArray = null;
if (!params.input1.isInSparseFormat())
inputArray = params.input1.getDenseBlock();
double [] outputArray = null;
if (!params.output.isInSparseFormat())
outputArray = params.output.getDenseBlock();
+ else
+ throw new DMLRuntimeException("Expected the output to be allocated in dense format");
long tmpNNZ = 0;
for (int p = 0; p < params.P; p++) {
