Java examples broken with mxnet mkldnn build

[arcadiaphy]

Description

I've built the scala-package with mxnet mkldnn and run the java demo ImageClassification, but the demo is broken on mobillenet.

Environment info (Required)

1. Java version: 1.8.0
2. Maven version: 3.6.0

Build info (Required if built from source)

mac build with clang

latest commit:
cab1dfa

Error Message:

Exception in thread "main" org.apache.mxnet.MXNetError: [01:45:18] src/ndarray/ndarray.cc:634: Check failed: !is_view: 
Stack trace:
  [bt] (0) 1   libmxnet.so                         0x0000000117c23f59 dmlc::LogMessageFatal::~LogMessageFatal() + 57
  [bt] (1) 2   libmxnet.so                         0x00000001195bbb3b mxnet::NDArray::GetMKLDNNData() const + 571
  [bt] (2) 3   libmxnet.so                         0x00000001195d5b8b void mxnet::CopyFromToDnsImpl<mshadow::cpu, mshadow::cpu>(mxnet::NDArray const&, mxnet::NDArray const&, mxnet::RunContext) + 2747
  [bt] (3) 4   libmxnet.so                         0x00000001195d4dca void mxnet::CopyFromToImpl<mshadow::cpu, mshadow::cpu>(mxnet::NDArray const&, mxnet::NDArray const&, mxnet::RunContext, std::__1::vector<mxnet::Resource, std::__1::allocator<mxnet::Resource> > const&) + 2458
  [bt] (4) 5   libmxnet.so                         0x00000001195d43e1 std::__1::__function::__func<mxnet::CopyFromTo(mxnet::NDArray const&, mxnet::NDArray const&, int, bool)::$_7, std::__1::allocator<mxnet::CopyFromTo(mxnet::NDArray const&, mxnet::NDArray const&, int, bool)::$_7>, void (mxnet::RunContext, mxnet::engine::CallbackOnComplete)>::operator()(mxnet::RunContext&&, mxnet::engine::CallbackOnComplete&&) + 65
  [bt] (5) 6   libmxnet.so                         0x000000011945af7d mxnet::engine::ThreadedEngine::ExecuteOprBlock(mxnet::RunContext, mxnet::engine::OprBlock*) + 701
  [bt] (6) 7   libmxnet.so                         0x000000011945f10e mxnet::engine::ThreadedEnginePerDevice::PushToExecute(mxnet::engine::OprBlock*, bool)::'lambda'()::operator()() const::'lambda'(std::__1::shared_ptr<dmlc::ManualEvent>)::operator()(std::__1::shared_ptr<dmlc::ManualEvent>) const + 190
  [bt] (7) 8   libmxnet.so                         0x000000011945efa9 std::__1::__function::__func<mxnet::engine::ThreadedEnginePerDevice::PushToExecute(mxnet::engine::OprBlock*, bool)::'lambda'()::operator()() const::'lambda'(std::__1::shared_ptr<dmlc::ManualEvent>), std::__1::allocator<mxnet::engine::ThreadedEnginePerDevice::PushToExecute(mxnet::engine::OprBlock*, bool)::'lambda'()::operator()() const::'lambda'(std::__1::shared_ptr<dmlc::ManualEvent>)>, void (std::__1::shared_ptr<dmlc::ManualEvent>)>::operator()(std::__1::shared_ptr<dmlc::ManualEvent>&&) + 41
  [bt] (8) 9   libmxnet.so                         0x000000011945c083 void* std::__1::__thread_proxy<std::__1::tuple<std::__1::unique_ptr<std::__1::__thread_struct, std::__1::default_delete<std::__1::__thread_struct> >, std::__1::function<void (std::__1::shared_ptr<dmlc::ManualEvent>)>, std::__1::shared_ptr<dmlc::ManualEvent> > >(void*) + 83


	at org.apache.mxnet.Base$.checkCall(Base.scala:111)
	at org.apache.mxnet.NDArray.internal(NDArray.scala:1231)
	at org.apache.mxnet.NDArray.toArray(NDArray.scala:1216)
	at org.apache.mxnet.infer.Predictor$$anonfun$11.apply(Predictor.scala:190)
	at org.apache.mxnet.infer.Predictor$$anonfun$11.apply(Predictor.scala:190)
	at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
	at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
	at scala.collection.mutable.ResizableArray$class.foreach(ResizableArray.scala:59)
	at scala.collection.mutable.ArrayBuffer.foreach(ArrayBuffer.scala:48)
	at scala.collection.TraversableLike$class.map(TraversableLike.scala:234)
	at scala.collection.AbstractTraversable.map(Traversable.scala:104)
	at org.apache.mxnet.infer.Predictor.predictImpl(Predictor.scala:190)
	at org.apache.mxnet.infer.Predictor.predict(Predictor.scala:156)
	at org.apache.mxnet.infer.javaapi.Predictor.predict(Predictor.scala:74)
	at mxnet.ImageClassification.main(ImageClassification.java:102)

Minimum reproducible example

https://github.com/apache/incubator-mxnet/blob/master/scala-package/mxnet-demo/java-demo/src/main/java/mxnet/ImageClassification.java

I've replaced the resnet-18 in the script with mobilenet:
model.zip

[mxnet-label-bot]

Hey, this is the MXNet Label Bot.
Thank you for submitting the issue! I will try and suggest some labels so that the appropriate MXNet community members can help resolve it.
Here are my recommended labels: Build

[leleamol]

@mxnet-label-bot add [Java, Scala, MKLDNN]

[arcadiaphy]

I've separated the bug from java interface, it comes from the slice operation in java predict:

  def predict(batch: DataBatch): IndexedSeq[NDArray] = {
    require(binded && paramsInitialized, "bind() and initParams() must be called first.")
    forward(batch, isTrain = Option(false))
    val pad = batch.pad
    getOutputsMerged().map(out => {
      val withoutPadding = out.slice(0, out.shape(0)-pad)
      val copied = withoutPadding.copy()
      withoutPadding.dispose()
      copied
    })
  }

I reproduced the bug with python code:

import mxnet as mx

# get symbol from gluon model
net = mx.gluon.model_zoo.vision.get_model('mobilenetv2_0.25')
net.hybridize()
net.initialize()
x = mx.nd.uniform(shape=(1, 3, 224, 224))
y = net(x)
sym = net._cached_graph[1]

exc = sym.simple_bind(ctx=mx.cpu(), grad_req='null', data=(1, 3, 224, 224))
y = exc.forward()[0]
# trigger the bug
print y.slice(0, 1)

I tested the above script on multiple mxnet versions: it works fine in 1.4, but triggers the error on 1.5.

Traceback (most recent call last):
  File "bug.py", line 14, in <module>
    print y.slice(0, 1)
  File "/Users/arcadia/work/mxnet-src/python/mxnet/ndarray/ndarray.py", line 194, in __repr__
    return '\n%s\n<%s %s @%s>' % (str(self.asnumpy()),
  File "/Users/arcadia/work/mxnet-src/python/mxnet/ndarray/ndarray.py", line 1996, in asnumpy
    ctypes.c_size_t(data.size)))
  File "/Users/arcadia/work/mxnet-src/python/mxnet/base.py", line 253, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: std::exception

[pengzhao-intel]

@TaoLv could you help to take a look for slice?

[TaoLv]

Sure. I will work on this later.

[TaoLv]

@arcadiaphy @pengzhao-intel This issue is root caused. The flatten layer before slice is not properly handled. It can be reproduced as below. We already have a fix in local and will submit a PR soon. Do you think we should have the fix into 1.5.0 release?

import mxnet as mx
import numpy as np
from mxnet import Context

np.random.seed(12345)

data = mx.symbol.Variable('data')
weight = mx.symbol.Variable('weight')
bias = mx.symbol.Variable('bias')
conv1= mx.symbol.Convolution(data = data, weight=weight, bias=bias, name='conv1', num_filter=64, kernel=(3,3), stride=(1,1))
flatten1 = mx.symbol.flatten(data = conv1)
slice1 = mx.symbol.slice(data = flatten1, begin=0, end=1)

shape = (2, 16, 224, 224)
val = np.random.rand(2, 16, 224, 224).astype(np.float32)
exe = slice1.simple_bind(Context.default_ctx, data=shape)
exe.arg_arrays[0][:] = val
exe.arg_arrays[1][:] = np.random.normal(size=exe.arg_arrays[1].shape)
p = exe.forward(is_train=False)
p[0].wait_to_read()
print(p[0])

[pengzhao-intel]

Fixed and closing now. Thanks to reporting the issue :)