Runtime Exception - Failed to read, graphDef and prefix cannot be null

[dreistheman]

@szaza I'm trying to run your Java API example (https://sites.google.com/view/tensorflow-example-java-api/home?authuser=0).
-Downloaded the graph and label file from the provided gdrive link, placed it in tensorflow-example-java\src\main\resources\YOLO
-edited the configuration file with the absolute paths of the yolo-voc.pb and yolo-voc-labels.pb

Unfortunately resulted to ServiceException and IllegalArgumentException. I'm trying to test this out so I can use it with my own trained model (at 20,000 steps now) following your tutorial too (Train YOLO with DarkFlow)

I am using Java JDK 1.8 and TensorFlow 1.9. Thank you in advance. :D

[szaza]

Hi,
Could you please copy the whole stack trace? It would help me a lot to see where the exception is thrown. Could you try to use relative paths in your configuration file? They should be relative from the /resources/YOLO directory. Please also try to start the server in debug mode and put brake points before the line where the exception is thrown. Run the application line by line from that point and hopefully you'll see the problem.

[dreistheman]

Hello, thank you for replying. :) I now changed the paths to relative from the Config.java's directory.

Here's the stack trace:

00:51:19.004 [main] ERROR edu.ml.tensorflow.util.IOUtil - Failed to read [../../../../resources/YOLO/yolo-voc.pb]!
00:51:19.004 [main] ERROR edu.ml.tensorflow.ObjectDetector - Download one of my graph file to run the program!
You can find my graphs here: https://drive.google.com/open?id=1GfS1Yle7Xari1tRUEi2EDYedFteAOaoN
2019-04-09 00:51:21.548448: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
Exception in thread "main" java.lang.IllegalArgumentException: graphDef and prefix cannot be null
at org.tensorflow.Graph.importGraphDef(Graph.java:127)
at org.tensorflow.Graph.importGraphDef(Graph.java:114)
at edu.ml.tensorflow.ObjectDetector.executeYOLOGraph(ObjectDetector.java:89)
at edu.ml.tensorflow.ObjectDetector.detect(ObjectDetector.java:49)
at edu.ml.tensorflow.Main.main(Main.java:8)

From line 89 in ObjectDetector.java

private float[] executeYOLOGraph(final Tensor<Float> image) {
    try (Graph graph = new Graph()) {
        graph.importGraphDef(GRAPH_DEF); //this line
        try (Session s = new Session(graph);
            Tensor<Float> result = s.runner().feed("input", image).fetch("output").run().get(0).expect(Float.class)) {
            float[] outputTensor = new float[YOLOClassifier.getInstance().getOutputSizeByShape(result)];
            FloatBuffer floatBuffer = FloatBuffer.wrap(outputTensor);
            result.writeTo(floatBuffer);
            return outputTensor;
        }
    }
}

The Config.java:

package edu.ml.tensorflow;

public interface Config {
    String GRAPH_FILE = "../../../../resources/YOLO/yolo-voc.pb";
    String LABEL_FILE = "../../../../resources/YOLO/yolo-voc-labels.txt";

    // Params used for image processing
    int SIZE = 416;
    float MEAN = 255f;

    // Output directory
    String OUTPUT_DIR = "../../../../../../sample";
}

[szaza]

I can see the problem, your paths are wrong.
String GRAPH_FILE = "../../../../resources/YOLO/yolo-voc.pb"; - it means that you have near the project root a /resources/YOLO directory with yolo-voc.pb.
I would recommend just to use the default values and place your .pb and txt file into the /src/main/resources folder. I would also recommend to use the ./sample value as OUTPUT_DIR;

tensorflow-example-java/

• src/
  • main/
    • resources/
      -> your .pb and .txt files should be placed here

Please read more detail about the maven project structure here: https://maven.apache.org/guides/introduction/introduction-to-the-standard-directory-layout.html; Gradle uses the same archetypes.

[dreistheman]

It worked now! Yes I actually read back again on your first reply mentioning "They should be relative from the /resources/YOLO directory." after just posting my recent response. My lack of experience assumed I would use the same convention of "navigating from the current directory".

Last quick small side question: I'm training 4 classes now at 20,000 steps on top of tiny-yolo-voc. Would it be better to stop at 30 or 40,000 for 4 classes? I generated 15000 image data from 37 images consisting those 4 classes.

[szaza]

It can happen, however it also depends on your training set and your settings. I would suggest to save your models after every 10,000 steps. When you discover that the confidence rate doesn't grow higher (it is constant) or it is lower than the steps before, than you definitely have to stop the training process.
I'm not sure whether you are familiar with Courseras Machine Learning courses: https://www.coursera.org/learn/machine-learning?
If your're not I would suggest to go through on the lessons hold by Andrew Ng. I've got many-many good answers related with my ML related questions.
Best regards, and good luck for your work!

[dreistheman]

Didn't change anything on the tiny-yolo-voc-4c.cfg except the "classes" and "filters" attribute. Is the confidence rate the "loss" or "moving ave loss", which in my case is playing around 0.1-0.4 value for a while now I noticed. Steps, loss, and moving ave loss are the only data displayed in the console training on darkflow.

I'm not familiar with that yet so I'll definitely have the course. Thank you for your outstanding help in all these. :)