omr-research / measuredetector Goto Github PK

Thanks for this efficient and effective measure detection model!

My collaborator and I were wondering if the trained embedding model could be shared as-is (section 4.2 of https://publik.tuwien.ac.at/files/publik_282966.pdf) either here or directly. If not, pointers to the 81k measure annotations (section 3) would help us train our own; https://github.com/apacha/OMR-Datasets#muscima-measure-annotations only gives 2.8k system measure annotations.

(Context: we've developed a low-resource neural alignment method and hope to apply it to audio-measure alignment, a task done manually in https://ismir2022program.ismir.net/lbd_399.html. Our method works best with semantic feature embeddings.)

Let us know!

The Readme links to https://github.com/waleedka/cocoapi, which doesn't exist anymore.

Is GPU mandatory to make this program work?
If yes, what are the hardware compatibility with your program?

Thanks!

I'm not very familiar with machine learning, so I haven't managed to successfully train any models myself. But the provided model from the demo works fine, and I'm using that for a while now.

However, I'm finding myself in the need of proper staff detection. I managed to work around this for a while by merging the detected bars, but it ultimately has some limits. I'd thus really appreciate it if you provided a trained model for this as well.

Hi,

I could install the package and dependencies.
I am running now the training based on the MUSCIMA++ dataset you provided.
As a test, I am running it now on a Linode Linux server with 2 CPUs and 24 GB or Ram.

As I am not using GPUs, I expected it takes some time. However the training has been running already from a few days (3 or 4). Do you think this is a normal situation?
It may be difficult to compare, but how much time do you usually need to run this training?
As there are only 140 pages in the MUSCIMA++ data, I was wondering is something is going wrong

Thank you

Hi,
Is there any running example or documentation available for the Measure Detector running with TF on java? I am currently trying to include the Measure Detector in one of my applications and I saw some old documentation on one of the commits like 3 years ago, so maybe there are still some examples.

I would assume that something like that should work:

       try(Graph graph = new Graph()) {
           graph.importGraphDef(Files.readAllBytes(modelPath) );
           // Opens up the session
           try (Session ses = new Session(graph)) {
               BufferedImage img = null;
               try {
                   img = ImageIO.read(new File("build/res/testTitle/split/test.png"));
                   System.out.println(img);
               } catch (IOException e) {
                   e.printStackTrace();
               }
               long t = System.currentTimeMillis();
              inputTensor = makeImageTensor(img);
                  output = ses.runner()
                      .feed("image_tensor", inputTensor)
                      .fetch("num_detections")
                      .fetch("detection_boxes")
                      .fetch("detection_scores")
                      .fetch("detection_classes")
                      .run();

But I always get the same result:

UINT8 tensor with shape [1, 761, 502, 3] [FLOAT tensor with shape [1], FLOAT tensor with shape [1, 600, 4], FLOAT tensor with shape [1, 600], FLOAT tensor with shape [1, 600]]

Help would be appreciated :)

Best Regards,
Lukas

I've installed the module using pip install --user omrdatasettools. When I run python prepare_muscima-pp_dataset.py in the StaveDetector folder, I then get:

Traceback (most recent call last):
  File "prepare_muscima-pp_dataset.py", line 4, in <module>
    from omrdatasettools.converters.ImageColorInverter import ImageColorInverter
ModuleNotFoundError: No module named 'omrdatasettools.converters'

While the requirements.txt as the tensorflow library pinned to version 1.13.1, this is of little use because pip doesn't provide any older versions for download. I also tried downgrading some packages, but this would require downgrading my whole system to Python 3.7, which is not feasible.

Hi
I can run the inference on the demo model without any issue.

However I tried to go to the steps of the training to export a "saved_model"

I went throught all your steps and I could train successfully based the dataset MUSCIMA++
and export the "saved_model".

However, if I try to run the inference script then I get an issue as follow:

Traceback (most recent call last):
File "standalone_inference_over_image.py", line 70, in
detection_graph = load_detection_graph(args.detection_inference_graph)
File "standalone_inference_over_image.py", line 47, in load_detection_graph
od_graph_def.ParseFromString(serialized_graph)
google.protobuf.message.DecodeError: Error parsing message with type 'tensorflow.GraphDef'

I tried several time and trying with different config files that you provide but I get the same issue in all the cases.

Do you have any idea on how to solve this issue?

Thank you very much