[incubator-mxnet] branch master updated: Enable large tensor inputs for Deconvolution layer (#19626)

[zh...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

zha0q1 pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git


The following commit(s) were added to refs/heads/master by this push:
     new 1f07693  Enable large tensor inputs for Deconvolution layer (#19626)
1f07693 is described below

commit 1f076931f537a43881004c95718f26bace83d192
Author: Rohit Kumar Srivastava <sr...@osu.edu>
AuthorDate: Tue Dec 8 18:23:29 2020 -0800

    Enable large tensor inputs for Deconvolution layer (#19626)
    
    Co-authored-by: Rohit Kumar Srivastava <sr...@buckeyemail.osu.edu>
---
 src/operator/nn/deconvolution-inl.h  |  2 +-
 tests/nightly/test_np_large_array.py | 22 ++++++++++++++++++++++
 2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/src/operator/nn/deconvolution-inl.h b/src/operator/nn/deconvolution-inl.h
index 0aac39a..f1e684e 100644
--- a/src/operator/nn/deconvolution-inl.h
+++ b/src/operator/nn/deconvolution-inl.h
@@ -454,7 +454,7 @@ class DeconvolutionOp {
  private:
   inline index_t InitTemp(const mshadow::Shape<4> &ishape,
                           const mshadow::Shape<4> &oshape) {
-    const int ksize = param_.kernel.Size();
+    const index_t ksize = param_.kernel.Size();
     shape_colunit_ = mshadow::Shape2(ishape[1] * ksize,
                                      oshape[2] * oshape[3]);
     shape_dstunit_ = mshadow::Shape3(param_.num_group,
diff --git a/tests/nightly/test_np_large_array.py b/tests/nightly/test_np_large_array.py
index b37c401..19ed36c 100644
--- a/tests/nightly/test_np_large_array.py
+++ b/tests/nightly/test_np_large_array.py
@@ -2372,3 +2372,25 @@ def test_convolution():
     assert inp.grad.shape == inp.shape
     assert inp.grad[0][0][0][0] == 0
 
+
+@use_np
+def test_deconvolution():
+    dim = 2
+    batch_size = 1
+    channel = 3
+    height = SMALL_Y
+    width = LARGE_X // 5
+    num_filter = 4
+    kernel = (4,) * dim   # => shape = (3, 3)
+
+    inp=mx.np.ones(shape=(batch_size, channel, 1, width))
+    weight = mx.np.ones(shape=(channel, num_filter, kernel[0], kernel[1]))
+    bias = mx.np.array(num_filter,)
+    inp.attach_grad()
+    with mx.autograd.record():
+        out = mx.npx.deconvolution(data=inp, weight=weight, num_filter=num_filter, \
+                                   kernel=kernel, no_bias=True)
+    assert out.shape == (batch_size, channel + 1, kernel[0], width + 3)
+    assert out[0][0][kernel[0]//2][kernel[1]] == channel * num_filter
+    assert inp.grad.shape == inp.shape
+    assert inp.grad[0][0][0][0] == 0