[Flaky Test] Fix test_gluon_model_zoo.test_models when MXNET_MKLDNN_D…

…EBUG=1 (apache#12069)

* reorder inputs

* use function flatten vs build in method

* update similar array atoi to 0.01

* fix reorder

* enable MXNET_MKLDNN_DEBUG in CI

* add exclude debug flag

* fix lint

* add warning log for excluded op

* retrigger

ci/docker/runtime_functions.sh

@@ -581,9 +581,7 @@ sanity_check() {
 unittest_ubuntu_python2_cpu() {
     set -ex
     export PYTHONPATH=./python/
-    # MXNET_MKLDNN_DEBUG is buggy and produces false positives
-    # https://github.com/apache/incubator-mxnet/issues/10026
-    #export MXNET_MKLDNN_DEBUG=1  # Ignored if not present
+    export MXNET_MKLDNN_DEBUG=1
     export MXNET_STORAGE_FALLBACK_LOG_VERBOSE=0
     nosetests-2.7 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_unittest.xml --verbose tests/python/unittest
     nosetests-2.7 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_train.xml --verbose tests/python/train
@@ -593,9 +591,7 @@ unittest_ubuntu_python2_cpu() {
 unittest_ubuntu_python3_cpu() {
     set -ex
     export PYTHONPATH=./python/
-    # MXNET_MKLDNN_DEBUG is buggy and produces false positives
-    # https://github.com/apache/incubator-mxnet/issues/10026
-    #export MXNET_MKLDNN_DEBUG=1  # Ignored if not present
+    export MXNET_MKLDNN_DEBUG=1  # Ignored if not present
     export MXNET_STORAGE_FALLBACK_LOG_VERBOSE=0
     nosetests-3.4 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_unittest.xml --verbose tests/python/unittest
     nosetests-3.4 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_quantization.xml --verbose tests/python/quantization
@@ -604,9 +600,7 @@ unittest_ubuntu_python3_cpu() {
 unittest_ubuntu_python3_cpu_mkldnn() {
     set -ex
     export PYTHONPATH=./python/
-    # MXNET_MKLDNN_DEBUG is buggy and produces false positives
-    # https://github.com/apache/incubator-mxnet/issues/10026
-    #export MXNET_MKLDNN_DEBUG=1  # Ignored if not present
+    export MXNET_MKLDNN_DEBUG=1  # Ignored if not present
     export MXNET_STORAGE_FALLBACK_LOG_VERBOSE=0
     nosetests-3.4 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_unittest.xml --verbose tests/python/unittest
     nosetests-3.4 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_mkl.xml --verbose tests/python/mkl
@@ -615,9 +609,7 @@ unittest_ubuntu_python3_cpu_mkldnn() {
 unittest_ubuntu_python2_gpu() {
     set -ex
     export PYTHONPATH=./python/
-    # MXNET_MKLDNN_DEBUG is buggy and produces false positives
-    # https://github.com/apache/incubator-mxnet/issues/10026
-    #export MXNET_MKLDNN_DEBUG=1  # Ignored if not present
+    export MXNET_MKLDNN_DEBUG=1  # Ignored if not present
     export MXNET_STORAGE_FALLBACK_LOG_VERBOSE=0
     nosetests-2.7 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_gpu.xml --verbose tests/python/gpu
 }
@@ -649,9 +641,7 @@ tutorialtest_ubuntu_python2_gpu() {
 unittest_ubuntu_python3_gpu() {
     set -ex
     export PYTHONPATH=./python/
-    # MXNET_MKLDNN_DEBUG is buggy and produces false positives
-    # https://github.com/apache/incubator-mxnet/issues/10026
-    #export MXNET_MKLDNN_DEBUG=1 # Ignored if not present
+    export MXNET_MKLDNN_DEBUG=1 # Ignored if not present
     export MXNET_STORAGE_FALLBACK_LOG_VERBOSE=0
     nosetests-3.4 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_gpu.xml --verbose tests/python/gpu
 }
@@ -678,9 +668,7 @@ unittest_ubuntu_tensorrt_gpu() {
 unittest_ubuntu_python2_quantization_gpu() {
     set -ex
     export PYTHONPATH=./python/
-    # MXNET_MKLDNN_DEBUG is buggy and produces false positives
-    # https://github.com/apache/incubator-mxnet/issues/10026
-    #export MXNET_MKLDNN_DEBUG=1  # Ignored if not present
+    export MXNET_MKLDNN_DEBUG=1  # Ignored if not present
     export MXNET_STORAGE_FALLBACK_LOG_VERBOSE=0
     nosetests-2.7 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_quantization_gpu.xml --verbose tests/python/quantization_gpu
 }
@@ -690,9 +678,7 @@ unittest_ubuntu_python2_quantization_gpu() {
 unittest_ubuntu_python3_quantization_gpu() {
     set -ex
     export PYTHONPATH=./python/
-    # MXNET_MKLDNN_DEBUG is buggy and produces false positives
-    # https://github.com/apache/incubator-mxnet/issues/10026
-    #export MXNET_MKLDNN_DEBUG=1 # Ignored if not present
+    export MXNET_MKLDNN_DEBUG=1 # Ignored if not present
     export MXNET_STORAGE_FALLBACK_LOG_VERBOSE=0
     nosetests-3.4 $NOSE_COVERAGE_ARGUMENTS --with-xunit --xunit-file nosetests_quantization_gpu.xml --verbose tests/python/quantization_gpu
 }

python/mxnet/gluon/nn/basic_layers.py

@@ -427,7 +427,7 @@ def __init__(self, **kwargs):
         super(Flatten, self).__init__(**kwargs)
 
     def hybrid_forward(self, F, x):
-        return x.reshape((0, -1))
+        return F.Flatten(x)
 
     def __repr__(self):
         return self.__class__.__name__

src/operator/nn/lrn.cc

@@ -204,6 +204,8 @@ NNVM_REGISTER_OP(_backward_LRN)
 .set_attr<nnvm::TIsBackward>("TIsBackward", true)
 #if MXNET_USE_MKLDNN == 1
 .set_attr<FComputeEx>("FComputeEx<cpu>", LRNGradComputeExCPU)
+// Native compute requires norm while MKLDNN does not so cannot be compared in debug mode
+.set_attr<bool>("TExcludeMKLDNNDebug", true)
 #endif
 .set_attr<FCompute>("FCompute<cpu>", LRNGradCompute<cpu>);
 

src/operator/nn/mkldnn/mkldnn_base.cc

@@ -473,9 +473,11 @@ void OpCheck::Init(const std::vector<mxnet::NDArray> &inputs_,
   auto ctx = inputs_[0].ctx();
   CHECK(!MKLDNNStream::Get()->HasOps());
   for (size_t i = 0; i < inputs_.size(); i++) {
-    inputs.emplace_back(inputs_[i].shape(), ctx,
-                        false, inputs_[i].dtype());
-    auto mem = inputs_[i].GetMKLDNNData();
+    NDArray data = inputs_[i];
+    inputs.emplace_back(data.shape(), ctx, false, data.dtype());
+    if (data.IsMKLDNNData() && data.IsView())
+        data = data.Reorder2Default();
+    auto mem = data.GetMKLDNNData();
     inputs[i].CopyFrom(*mem);
   }
   for (size_t i = 0; i < outputs_.size(); i++) {
@@ -494,6 +496,11 @@ void OpCheck::Run(mxnet::FCompute fn, const nnvm::NodeAttrs &attrs,
                   const std::vector<mxnet::NDArray> &inputs_,
                   const std::vector<mxnet::OpReqType> &req,
                   const std::vector<mxnet::NDArray> &outputs_) {
+  static auto& is_excluded = Op::GetAttr<bool>("TExcludeMKLDNNDebug");
+  if (is_excluded.get(attrs.op, false)) {
+    LOG(WARNING) << attrs.op->name << " not checked. TExcludeMKLDNNDebug flag present";
+    return;
+  }
   std::vector<mxnet::TBlob> in_blobs(inputs.size());
   for (size_t i = 0; i < in_blobs.size(); i++) in_blobs[i] = inputs[i].data();
   std::vector<mxnet::TBlob> out_blobs(outputs.size());
@@ -509,7 +516,7 @@ void OpCheck::Run(mxnet::FCompute fn, const nnvm::NodeAttrs &attrs,
     if (req[i] == kNullOp)
       continue;
     MSHADOW_TYPE_SWITCH(outputs[i].dtype(), DType, {
-      bool similar = SimilarArray<DType>(outputs[i], outputs_[i], 1e-3, 1e-3);
+      bool similar = SimilarArray<DType>(outputs[i], outputs_[i], 1e-2, 1e-2);
       if (!similar) {
         LOG(ERROR) << attrs.op->name << " fails";
       }