Mismatch between shape Java API

[androuino]

Hi, I’ve finished training the yolo3_darknet53 params and wanted to test it with Java API, however am having this error which I have no idea where it coming from:

 Exception in thread “main” java.lang.IllegalArgumentException: requirement failed:
 Mismatch between shape (100,1) and (1,100,4)
    	at scala.Predef$.require(Predef.scala:224)
    	at org.apache.mxnet.NDArray$$anonfun$9.apply(NDArray.scala:448)
    	at org.apache.mxnet.NDArray$$anonfun$9.apply(NDArray.scala:445)
    	at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
    	at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
    	at scala.collection.Iterator$class.foreach(Iterator.scala:893)
    	at scala.collection.AbstractIterator.foreach(Iterator.scala:1336)
    	at scala.collection.IterableLike$class.foreach(IterableLike.scala:72)
    	at scala.collection.AbstractIterable.foreach(Iterable.scala:54)
    	at scala.collection.TraversableLike$class.map(TraversableLike.scala:234)
    	at scala.collection.AbstractTraversable.map(Traversable.scala:104)
    	at org.apache.mxnet.NDArray$.concatenate(NDArray.scala:445)
    	at org.apache.mxnet.module.BaseModule$$anonfun$2.apply(BaseModule.scala:267)
    	at org.apache.mxnet.module.BaseModule$$anonfun$2.apply(BaseModule.scala:267)
    	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.module.BaseModule.predict(BaseModule.scala:267)
    	at org.apache.mxnet.infer.Predictor$$anonfun$11.apply(Predictor.scala:210)
    	at org.apache.mxnet.infer.Predictor$$anonfun$11.apply(Predictor.scala:210)
    	at org.apache.mxnet.infer.MXNetThreadPoolHandler$$anon$4.call(MXNetHandler.scala:73)
    	at java.util.concurrent.FutureTask.run(FutureTask.java:266)
    	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
    	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
    	at java.lang.Thread.run(Thread.java:748)
FAILURE: Build failed with an exception.

* What went wrong:
Execution failed for task ':ObjectDetectionTutorial.main()'.
> Process 'command '/usr/lib/jvm/java-8-openjdk-amd64/bin/java'' finished with non-zero exit value 1

* Try:
Run with --stacktrace option to get the stack trace. Run with --info or --debug option to get more log output. Run with --scan to get full insights.

* Get more help at https://help.gradle.org

BUILD FAILED in 1s
2 actionable tasks: 2 executed
Process 'command '/usr/lib/jvm/java-8-openjdk-amd64/bin/java'' finished with non-zero exit value 1
10:16:24: Task execution finished 'ObjectDetectionTutorial.main()'.

I only have 1 class and my input image is 512 and pretrained_base=false. Any help would be much appreciated. Thank you in advance.

[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: Doc

[frankfliu]

@mxnet-label-bot add [scala, bug]

[lanking520]

@androuino It's a general shape mismatch problem sounds like.

What is the input shape of your model? Are you trying to do object detection on it?

It will be better if you can share me your model and I can have a quick evaluation on it.

[androuino]

Thanks @lanking520 for the response. Yes am trying to do object detection and the input shape of my model is [1, 3, 512, 512] (if this is what you're looking for). This is my symbol.json file as you have requested.
https://gist.github.com/androuino/0b5e36c2fb39be2bda6cd8fc8bdadd05
and this is my params
https://filebin.net/n89gzmb285oxre2g
Thank you so much!.

[lanking520]

@androuino Which version of MXNet Java are you using?

[androuino]

Hi @lanking520, am using the latest mxnet version (1.4.0) both java library and for training the model in linux gpu machine. I actually trained a resnet50 model hoping that training is different from yolo but they are throwing the same error. Do you think by changing mxnet to previous version would fix the problem? Thanks for your response.

[androuino]

Tried the versions starting from 1.2.0 up to 1.3.1 but without success, all these versions has no ObjectDetector class, or maybe the implementation is different.

[lanking520]

@androuino Thank you, I can reproduce this problem in the most recent 1.5.0-SNAPSHOT. This problem is a bug in the concat operator in Scala it seemed. I will be working on that from today on to see how I can fix it.

[androuino]

Hi @lanking520 thank you for addressing this issue. I'm really looking forward for the fix. I hope it don't give you hard time fixing it. If only the java library is open source, my team could help fixing the issue as well. Anyway thank you so much, hoping to hear some good news about this issue soon.

[androuino]

I would also like to ask why am getting this error also:
https://gist.github.com/androuino/213131592b1d22763bd2a0f98fff7ed8
This happens when I use the java library: compile group: 'org.apache.mxnet', name: 'mxnet-full_2.11-linux-x86_64-gpu', version: '1.4.0'.
My machine is Ubuntu 18.04 with Cuda 9.0 and nvidia-390 installed for GeForce GTX1080ti.
I tried the same library with a fresh install of XUbuntu 16.04 with Cuda 9.2 and nvidia-396 installed for GeForce1070ti GPU but without success, am having the same error. Am using Intellij 2019.1 for testing. For Ubuntu 18.04, I installed the mxnet library through pip install mxnet-cu90, for XUbuntu 16.04, I installed the mxnet library from source (cmake). Thank you.

[lanking520]

@androuino Currently, we publish our CUDA package under 9.2 version, all version prior will not be supported unfortunatly. Please install CUDA 9.2 on your machine to get it work.

The reason that might cause 9.2 not working is the CUDA path not properly set. As an alternative, you can use the script we use to install cuda and cudnn. For your use case it will be:

setup_gpu_build_tools.sh cu92 /path-you-want-to-place-cuda

If you are interested, you can join our slack channel https://mxnet.apache.org/versions/master/community/contribute.html#slack where you will expect faster response.

[lanking520]

Minimum reproducible Java code:

import org.apache.mxnet.infer.javaapi.Predictor;
import org.apache.mxnet.javaapi.*;

import java.util.ArrayList;
import java.util.Arrays;
import java.util.List;

public class YoloInference {
    public static void main(String[] args) {
        // Prepare the predictor
        List<DataDesc> inputDesc = new ArrayList<>();
        List<Context> contexts = new ArrayList<>();
        DataDesc dataDesc = new DataDesc("data", new Shape(new int[]{1, 3, 512, 512}), DType.Float32(), "NCHW");
        Context context = Context.cpu();
        inputDesc.add(dataDesc);
        contexts.add(context);
        Predictor predictor = new Predictor("yolo/yolo3_darknet53", inputDesc, contexts, 0);
        // Prepare the data
        NDArray nd = NDArray.ones(Context.cpu(), new int[]{1, 3, 512, 512});
        // Do inference
        System.out.println(nd);
        predictor.predictWithNDArray(Arrays.asList(nd));
    }
}

Where as Python did not have this problem:

import mxnet as mx
from mxnet import nd

sym, arg_params, aux_params = mx.model.load_checkpoint(prefix='yolo3_darknet53', epoch=0)

mod = mx.mod.Module(symbol=sym,
                    data_names=['data'],
                    context=mx.cpu(),
                    label_names=None)
mod.bind(for_training=False,
         data_shapes=[('data', (1, 3, 512, 512))],
         label_shapes=mod._label_shapes)

mod.set_params(arg_params, aux_params, allow_missing=True)

sample_input = nd.ones((1, 3, 512, 512))

data_iter = mx.io.NDArrayIter(sample_input, None, 1)
result = mod.predict(data_iter)
for item in result:
    print(item.shape)

[lanking520]

I add a PR to fix this problem. The problem is the wrong ways of concatenating the NDArrays. It used to be:

  outputBatches = [
         [a1, a2, a3], // batch a
         [b1, b2, b3]  // batch b
   ]
  result = [
         NDArray, // [a1, a2, a3]
         NDArray, // [b1, b2, b3]
  ]

Now is:

  result = [
         NDArray, // [a1, b1]
         NDArray, // [a2, b2]
         NDArray, // [a3, b3]
  ]

[androuino]

Hi @lanking520, wow the fix is here already! I am now testing the fix for the Java API. I will also highly consider the script for installing CUDA in a fresh new installed Ubuntu OS.

I'll be happy to join the slack channel. Thanks again for all of this.

[androuino]

Hi @lanking520. After pulling your changes for the fix of my issue, I tried it right away but am having now a strange error using the example of ObjectDetection.java class but I made some alteration to it to meet my requirements. Now it looks like this:
https://gist.github.com/androuino/7808b6fdf05e3122a03f35c63d3a5f89

I followed the step by step tutorial here /~https://github.com/apache/incubator-mxnet/tree/master/scala-package/mxnet-demo/java-demo for running the demo.

and this is the error am having:

SLF4J: Failed to load class "org.slf4j.impl.StaticLoggerBinder". SLF4J: Defaulting to no-operation (NOP) logger implementation SLF4J: See http://www.slf4j.org/codes.html#StaticLoggerBinder for further details. [16:52:43] src/nnvm/legacy_json_util.cc:209: Loading symbol saved by previous version v1.4.0. Attempting to upgrade... [16:52:43] src/nnvm/legacy_json_util.cc:217: Symbol successfully upgraded! Exception in thread "main" java.lang.IllegalStateException: Invalid output shapes, expected: Vector(DataDesc[data,(1,3,512,512),float32,NCHW]).length, actual: ArrayBuffer((yolov30_slice_axis1_output,(1,100,1)), (yolov30_slice_axis2_output,(1,100,1)), (yolov30_slice_axis3_output,(1,100,4))).length. at org.apache.mxnet.infer.ObjectDetector.getImageClassifier(ObjectDetector.scala:202) at org.apache.mxnet.infer.ObjectDetector.<init>(ObjectDetector.scala:52) at org.apache.mxnet.infer.javaapi.ObjectDetector.<init>(ObjectDetector.scala:58) at mxnet.ObjectDetection.runObjectDetectionSingle(ObjectDetection.java:77) at mxnet.ObjectDetection.main(ObjectDetection.java:86)

The image I was used to test is 512x512 in size and the model that I've trained is also set to 512.
This is my train_yolo3.py script: https://gist.github.com/androuino/af5212923534a204b155c01b3bacb7f1
Then run the script with this command python train.py --gpus 0 --batch-size 2 --data-shape 512

If you want the files that am using including the test image, I could email it you. I couldn't upload it publicly so I preferred to email it directly to you.

Thanks for any enlightenment that you could give me with this error or at least tell me if I did something or missed something with the code: https://gist.github.com/androuino/7808b6fdf05e3122a03f35c63d3a5f89 .