You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@systemml.apache.org by du...@apache.org on 2017/04/12 21:49:46 UTC
incubator-systemml git commit: [SYSTEMML-1516] Improve output size
calculation in conv2d & max_pool2d
Repository: incubator-systemml
Updated Branches:
refs/heads/master fb82482b0 -> 4a941e44e
[SYSTEMML-1516] Improve output size calculation in conv2d & max_pool2d
The calculation of the output sizes (`Hout`, `Wout`) for conv2d and
max_pool2d layers rely on a flooring operation to only allow for valid
operations. Previously, we simply relied on the `as.integer` call to
implicitly perform this flooring operation during truncation of the
double to an integer. While this gave the correct results, it was less
readable than it should be, and relied on truncation behavior rather
than explicit mathematical behavior. This commit simply adds an
explicit `floor` operation to each calculation.
Closes #459.
Project: http://git-wip-us.apache.org/repos/asf/incubator-systemml/repo
Commit: http://git-wip-us.apache.org/repos/asf/incubator-systemml/commit/4a941e44
Tree: http://git-wip-us.apache.org/repos/asf/incubator-systemml/tree/4a941e44
Diff: http://git-wip-us.apache.org/repos/asf/incubator-systemml/diff/4a941e44
Branch: refs/heads/master
Commit: 4a941e44e008a9e65e6a5422d727c720980b0a3d
Parents: fb82482
Author: Mike Dusenberry <mw...@us.ibm.com>
Authored: Wed Apr 12 14:49:09 2017 -0700
Committer: Mike Dusenberry <mw...@us.ibm.com>
Committed: Wed Apr 12 14:49:09 2017 -0700
----------------------------------------------------------------------
scripts/staging/SystemML-NN/nn/layers/batch_norm1d.dml | 2 +-
scripts/staging/SystemML-NN/nn/layers/batch_norm2d.dml | 2 +-
scripts/staging/SystemML-NN/nn/layers/conv2d.dml | 4 ++--
.../staging/SystemML-NN/nn/layers/conv2d_builtin.dml | 4 ++--
scripts/staging/SystemML-NN/nn/layers/max_pool2d.dml | 4 ++--
.../SystemML-NN/nn/layers/max_pool2d_builtin.dml | 4 ++--
scripts/staging/SystemML-NN/nn/test/conv2d_simple.dml | 6 ++----
scripts/staging/SystemML-NN/nn/test/grad_check.dml | 12 ++++++------
.../staging/SystemML-NN/nn/test/max_pool2d_simple.dml | 4 ++--
scripts/staging/SystemML-NN/nn/test/test.dml | 9 +++++++--
scripts/staging/SystemML-NN/nn/util.dml | 8 ++++----
11 files changed, 31 insertions(+), 28 deletions(-)
----------------------------------------------------------------------
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/layers/batch_norm1d.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/layers/batch_norm1d.dml b/scripts/staging/SystemML-NN/nn/layers/batch_norm1d.dml
index 9ecbd77..2ccffdb 100644
--- a/scripts/staging/SystemML-NN/nn/layers/batch_norm1d.dml
+++ b/scripts/staging/SystemML-NN/nn/layers/batch_norm1d.dml
@@ -81,7 +81,7 @@ forward = function(matrix[double] X, matrix[double] gamma, matrix[double] beta,
*/
N = nrow(X)
- if(mode == 'train') {
+ if (mode == 'train') {
# Compute feature-wise mean and variance
mean = colMeans(X) # shape (1, D)
# var = (1/N) * colSums((X-mean)^2)
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/layers/batch_norm2d.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/layers/batch_norm2d.dml b/scripts/staging/SystemML-NN/nn/layers/batch_norm2d.dml
index fb25b2c..49c6746 100644
--- a/scripts/staging/SystemML-NN/nn/layers/batch_norm2d.dml
+++ b/scripts/staging/SystemML-NN/nn/layers/batch_norm2d.dml
@@ -88,7 +88,7 @@ forward = function(matrix[double] X, matrix[double] gamma, matrix[double] beta,
*/
N = nrow(X)
- if(mode == 'train') {
+ if (mode == 'train') {
# Compute channel-wise mean and variance
# Since we don't have tensors, we will compute the means and variances in a piece-wise fashion.
# - mean of total group is mean of subgroup means
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/layers/conv2d.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/layers/conv2d.dml b/scripts/staging/SystemML-NN/nn/layers/conv2d.dml
index 7aeec16..9d03568 100644
--- a/scripts/staging/SystemML-NN/nn/layers/conv2d.dml
+++ b/scripts/staging/SystemML-NN/nn/layers/conv2d.dml
@@ -67,8 +67,8 @@ forward = function(matrix[double] X, matrix[double] W, matrix[double] b,
*/
N = nrow(X)
F = nrow(W)
- Hout = as.integer((Hin + 2*padh - Hf)/strideh + 1)
- Wout = as.integer((Win + 2*padw - Wf)/stridew + 1)
+ Hout = as.integer(floor((Hin + 2*padh - Hf)/strideh + 1))
+ Wout = as.integer(floor((Win + 2*padw - Wf)/stridew + 1))
# Create output volume
out = matrix(0, rows=N, cols=F*Hout*Wout)
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/layers/conv2d_builtin.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/layers/conv2d_builtin.dml b/scripts/staging/SystemML-NN/nn/layers/conv2d_builtin.dml
index e7771ba..bda7a9c 100644
--- a/scripts/staging/SystemML-NN/nn/layers/conv2d_builtin.dml
+++ b/scripts/staging/SystemML-NN/nn/layers/conv2d_builtin.dml
@@ -66,8 +66,8 @@ forward = function(matrix[double] X, matrix[double] W, matrix[double] b,
*/
N = nrow(X)
F = nrow(W)
- Hout = as.integer((Hin + 2*padh - Hf)/strideh + 1)
- Wout = as.integer((Win + 2*padw - Wf)/stridew + 1)
+ Hout = as.integer(floor((Hin + 2*padh - Hf)/strideh + 1))
+ Wout = as.integer(floor((Win + 2*padw - Wf)/stridew + 1))
# Convolution - built-in implementation
out = conv2d(X, W, input_shape=[N,C,Hin,Win], filter_shape=[F,C,Hf,Wf],
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/layers/max_pool2d.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/layers/max_pool2d.dml b/scripts/staging/SystemML-NN/nn/layers/max_pool2d.dml
index ef1499a..fba1a4c 100644
--- a/scripts/staging/SystemML-NN/nn/layers/max_pool2d.dml
+++ b/scripts/staging/SystemML-NN/nn/layers/max_pool2d.dml
@@ -57,8 +57,8 @@ forward = function(matrix[double] X, int C, int Hin, int Win, int Hf, int Wf,
* - Wout: Output width.
*/
N = nrow(X)
- Hout = as.integer((Hin + 2*padh - Hf)/strideh + 1)
- Wout = as.integer((Win + 2*padw - Wf)/stridew + 1)
+ Hout = as.integer(floor((Hin + 2*padh - Hf)/strideh + 1))
+ Wout = as.integer(floor((Win + 2*padw - Wf)/stridew + 1))
pad_value = -1/0 # in max pooling we pad with -infinity
# Create output volume
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/layers/max_pool2d_builtin.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/layers/max_pool2d_builtin.dml b/scripts/staging/SystemML-NN/nn/layers/max_pool2d_builtin.dml
index 65ba71f..880f818 100644
--- a/scripts/staging/SystemML-NN/nn/layers/max_pool2d_builtin.dml
+++ b/scripts/staging/SystemML-NN/nn/layers/max_pool2d_builtin.dml
@@ -56,8 +56,8 @@ forward = function(matrix[double] X, int C, int Hin, int Win, int Hf, int Wf,
* - Wout: Output width.
*/
N = nrow(X)
- Hout = as.integer((Hin-Hf)/strideh + 1)
- Wout = as.integer((Win-Wf)/stridew + 1)
+ Hout = as.integer(floor((Hin + 2*padh - Hf)/strideh + 1))
+ Wout = as.integer(floor((Win + 2*padw - Wf)/stridew + 1))
# Max pooling - built-in implementation
out = max_pool(X, input_shape=[N,C,Hin,Win], pool_size=[Hf,Wf],
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/test/conv2d_simple.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/test/conv2d_simple.dml b/scripts/staging/SystemML-NN/nn/test/conv2d_simple.dml
index 05f0f7d..9f126d0 100644
--- a/scripts/staging/SystemML-NN/nn/test/conv2d_simple.dml
+++ b/scripts/staging/SystemML-NN/nn/test/conv2d_simple.dml
@@ -57,8 +57,8 @@ forward = function(matrix[double] X, matrix[double] W, matrix[double] b,
*/
N = nrow(X)
F = nrow(W)
- Hout = as.integer((Hin + 2*padh - Hf)/strideh + 1)
- Wout = as.integer((Win + 2*padw - Wf)/stridew + 1)
+ Hout = as.integer(floor((Hin + 2*padh - Hf)/strideh + 1))
+ Wout = as.integer(floor((Win + 2*padw - Wf)/stridew + 1))
# Create output volume
out = matrix(0, rows=N, cols=F*Hout*Wout)
@@ -131,8 +131,6 @@ backward = function(matrix[double] dout, int Hout, int Wout,
*/
N = nrow(X)
F = nrow(W)
- Hout = as.integer((Hin + 2*padh - Hf)/strideh + 1)
- Wout = as.integer((Win + 2*padw - Wf)/stridew + 1)
# Create gradient volumes
dX = matrix(0, rows=N, cols=C*Hin*Win)
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/test/grad_check.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/test/grad_check.dml b/scripts/staging/SystemML-NN/nn/test/grad_check.dml
index 516fe2a..f3bc9a7 100644
--- a/scripts/staging/SystemML-NN/nn/test/grad_check.dml
+++ b/scripts/staging/SystemML-NN/nn/test/grad_check.dml
@@ -1026,8 +1026,8 @@ max_pool2d = function() {
for (pad in 0:1) {
print(" - Grad checking w/ pad="+pad+".")
- Hout = as.integer((Hin + 2*pad - Hf)/stride + 1)
- Wout = as.integer((Win + 2*pad - Wf)/stride + 1)
+ Hout = as.integer(floor((Hin + 2*pad - Hf)/stride + 1))
+ Wout = as.integer(floor((Win + 2*pad - Wf)/stride + 1))
y = rand(rows=N, cols=C*Hout*Wout)
# Compute analytical gradients of loss wrt parameters
@@ -1075,8 +1075,8 @@ max_pool2d_builtin = function() {
for (pad in 0:1) {
print(" - Grad checking w/ pad="+pad+".")
- Hout = as.integer((Hin + 2 * pad - Hf) / stride + 1)
- Wout = as.integer((Win + 2 * pad - Wf) / stride + 1)
+ Hout = as.integer(floor((Hin + 2 * pad - Hf) / stride + 1))
+ Wout = as.integer(floor((Win + 2 * pad - Wf) / stride + 1))
y = rand(rows=N, cols=C*Hout*Wout)
# Compute analytical gradients of loss wrt parameters
@@ -1128,8 +1128,8 @@ max_pool2d_simple = function() {
for (pad in 0:1) {
print(" - Grad checking w/ pad="+pad+".")
- Hout = as.integer((Hin + 2*pad - Hf)/stride + 1)
- Wout = as.integer((Win + 2*pad - Wf)/stride + 1)
+ Hout = as.integer(floor((Hin + 2*pad - Hf)/stride + 1))
+ Wout = as.integer(floor((Win + 2*pad - Wf)/stride + 1))
y = rand(rows=N, cols=C*Hout*Wout)
# Compute analytical gradients of loss wrt parameters
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/test/max_pool2d_simple.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/test/max_pool2d_simple.dml b/scripts/staging/SystemML-NN/nn/test/max_pool2d_simple.dml
index dee1a48..188bd6e 100644
--- a/scripts/staging/SystemML-NN/nn/test/max_pool2d_simple.dml
+++ b/scripts/staging/SystemML-NN/nn/test/max_pool2d_simple.dml
@@ -55,8 +55,8 @@ forward = function(matrix[double] X, int C, int Hin, int Win, int Hf, int Wf,
* - Wout: Output width.
*/
N = nrow(X)
- Hout = as.integer((Hin + 2*padh - Hf)/strideh + 1)
- Wout = as.integer((Win + 2*padw - Wf)/stridew + 1)
+ Hout = as.integer(floor((Hin + 2*padh - Hf)/strideh + 1))
+ Wout = as.integer(floor((Win + 2*padw - Wf)/stridew + 1))
# Create output volume
out = matrix(0, rows=N, cols=C*Hout*Wout)
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/test/test.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/test/test.dml b/scripts/staging/SystemML-NN/nn/test/test.dml
index 3928fac..5a0390f 100644
--- a/scripts/staging/SystemML-NN/nn/test/test.dml
+++ b/scripts/staging/SystemML-NN/nn/test/test.dml
@@ -147,8 +147,8 @@ im2col = function() {
Wf = 3 # filter width
stride = 2
pad = (Hin * stride - Hin + Hf - stride) / 2
- Hout = as.integer((Hin + 2 * pad - Hf) / stride + 1)
- Wout = as.integer((Win + 2 * pad - Wf) / stride + 1)
+ Hout = as.integer(floor((Hin + 2*pad - Hf)/stride + 1))
+ Wout = as.integer(floor((Win + 2*pad - Wf)/stride + 1))
x = rand(rows=C, cols=Hin*Win)
# pad
@@ -157,6 +157,11 @@ im2col = function() {
# im2col
x_cols = util::im2col(x_pad, Hin+2*pad, Win+2*pad, Hf, Wf, stride, stride)
+ if (ncol(x_cols) != Hout*Wout) {
+ print("ERROR: im2col does not yield the correct output size: "
+ + ncol(x_cols)+" (actual) vs. "+Hout*Wout+" (correct).")
+ }
+
# col2im
x_pad2 = util::col2im(x_cols, C, Hin+2*pad, Win+2*pad, Hf, Wf, stride, stride, "none")
http://git-wip-us.apache.org/repos/asf/incubator-systemml/blob/4a941e44/scripts/staging/SystemML-NN/nn/util.dml
----------------------------------------------------------------------
diff --git a/scripts/staging/SystemML-NN/nn/util.dml b/scripts/staging/SystemML-NN/nn/util.dml
index 62a90f2..3a73f08 100644
--- a/scripts/staging/SystemML-NN/nn/util.dml
+++ b/scripts/staging/SystemML-NN/nn/util.dml
@@ -64,8 +64,8 @@ im2col = function(matrix[double] img, int Hin, int Win, int Hf, int Wf, int stri
* out into columns, of shape (C*Hf*Wf, Hout*Wout).
*/
C = nrow(img)
- Hout = as.integer((Hin-Hf)/strideh + 1)
- Wout = as.integer((Win-Wf)/stridew + 1)
+ Hout = as.integer(floor((Hin-Hf)/strideh + 1))
+ Wout = as.integer(floor((Win-Wf)/stridew + 1))
# Note: We start with `img_cols` transposed to allow for row-major
# left-indexing inside the loop, which is more performant.
@@ -119,8 +119,8 @@ col2im = function(matrix[double] img_cols, int C, int Hin, int Win, int Hf, int
* Outputs:
* - img: Input image, of shape (C, Hin*Win).
*/
- Hout = as.integer((Hin-Hf)/strideh + 1)
- Wout = as.integer((Win-Wf)/stridew + 1)
+ Hout = as.integer(floor((Hin-Hf)/strideh + 1))
+ Wout = as.integer(floor((Win-Wf)/stridew + 1))
img = matrix(0, rows=C, cols=Hin*Win) # zeros
for (hout in 1:Hout) { # all output rows