[incubator-mxnet] branch v1.4.x updated: Fix deconvolution / PR 13421 (#13433)

[an...@apache.org]

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

anirudh2290 pushed a commit to branch v1.4.x
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git


The following commit(s) were added to refs/heads/v1.4.x by this push:
     new 79532d9  Fix deconvolution  / PR 13421 (#13433)
79532d9 is described below

commit 79532d9cc176363eab1e1afbdcf7864131178e4d
Author: Alexander Zai <az...@gmail.com>
AuthorDate: Thu Nov 29 17:24:18 2018 -0800

    Fix deconvolution  / PR 13421 (#13433)
    
    * add test case
    
    * revert refactor
    
    * use with seed decorator
    
    * retrigger
    
    * remove seed
    
    * remove iteration
    
    * remove old test
    
    * update deconvolution test to have filter length that triggers mkldnn reorder
---
 src/operator/nn/mkldnn/mkldnn_deconvolution.cc |  6 ++++--
 tests/python/gpu/test_operator_gpu.py          | 10 +++++-----
 2 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/src/operator/nn/mkldnn/mkldnn_deconvolution.cc b/src/operator/nn/mkldnn/mkldnn_deconvolution.cc
index a6d6b24..4da48fa 100644
--- a/src/operator/nn/mkldnn/mkldnn_deconvolution.cc
+++ b/src/operator/nn/mkldnn/mkldnn_deconvolution.cc
@@ -262,12 +262,14 @@ void MKLDNNDeconvForward::SetDataHandle(const DeconvolutionParam& param,
     // For inference, we want to reorder the weight array so we don't need to
     // reorder data every time.
     if (weight.IsDefaultData()) {
+      weight_mem = GetWeights(weight, fwd_pd.weights_primitive_desc(), param.num_group);
       // We also need to modify the layout on the original weight array. The
       // data conversion happens after the weight array is used.
       const_cast<NDArray&>(weight).MKLDNNDataReorderAsync(fwd_pd.weights_primitive_desc());
+    } else {
+      weight_mem = weight.GetMKLDNNData();
+      CHECK(weight_mem->get_primitive_desc() == fwd_pd.weights_primitive_desc());
     }
-    weight_mem = weight.GetMKLDNNData();
-    CHECK(weight_mem->get_primitive_desc() == fwd_pd.weights_primitive_desc());
   }
   auto out_mem = CreateMKLDNNMem(out_data[deconv::kOut],
       fwd_pd.diff_src_primitive_desc(), req[deconv::kOut]);
diff --git a/tests/python/gpu/test_operator_gpu.py b/tests/python/gpu/test_operator_gpu.py
index 8054937..7a7c6f6 100644
--- a/tests/python/gpu/test_operator_gpu.py
+++ b/tests/python/gpu/test_operator_gpu.py
@@ -683,11 +683,11 @@ def test_deconvolution_options():
     check_consistency_NxM([sym, sym_no_cudnn], ctx_list)
 
     # 2D deconvolution
-    ctx_list = [{'ctx': mx.gpu(0), 'deconv_data': (2, 2, 10, 10), 'type_dict': {'deconv_data': np.float64}},
-                {'ctx': mx.gpu(0), 'deconv_data': (2, 2, 10, 10), 'type_dict': {'deconv_data': np.float32}},
-                {'ctx': mx.gpu(0), 'deconv_data': (2, 2, 10, 10), 'type_dict': {'deconv_data': np.float16}},
-                {'ctx': mx.cpu(0), 'deconv_data': (2, 2, 10, 10), 'type_dict': {'deconv_data': np.float64}},
-                {'ctx': mx.cpu(0), 'deconv_data': (2, 2, 10, 10), 'type_dict': {'deconv_data': np.float32}}]
+    ctx_list = [{'ctx': mx.gpu(0), 'deconv_data': (2, 8, 10, 10), 'type_dict': {'deconv_data': np.float64}},
+                {'ctx': mx.gpu(0), 'deconv_data': (2, 8, 10, 10), 'type_dict': {'deconv_data': np.float32}},
+                {'ctx': mx.gpu(0), 'deconv_data': (2, 8, 10, 10), 'type_dict': {'deconv_data': np.float16}},
+                {'ctx': mx.cpu(0), 'deconv_data': (2, 8, 10, 10), 'type_dict': {'deconv_data': np.float64}},
+                {'ctx': mx.cpu(0), 'deconv_data': (2, 8, 10, 10), 'type_dict': {'deconv_data': np.float32}}]
     # Pad > 0
     sym = mx.sym.Deconvolution(num_filter=2, kernel=(3,3), pad=(1,1), name='deconv')
     sym_no_cudnn = mx.sym.Deconvolution(num_filter=2, kernel=(3,3), pad=(1,1), cudnn_off=True, name='deconv')