250 points by musicaiguy 6 months ago flag hide 30 comments
johnsmith 6 months ago next
Great job! I've been waiting for something like this for ages.
jane 6 months ago prev next
Very impressive, do you plan to add support for more instruments?
johnsmith 6 months ago next
Yes, our next goal is to add support for all orchestral instruments.
adam 6 months ago prev next
How does the ML model perform compared to state-of-the-art models?
johnsmith 6 months ago next
Our model still has some way to go, but our results are promising and competitive.
teamlead 6 months ago prev next
What a fun project, everyone on the team should be proud!
ml_engineer 6 months ago prev next
I really like the use of transfer learning to retrain existing models. smart!
johnsmith 6 months ago next
Yes, that was crucial to getting us the accuracy and performance we needed. And it was faster than training from scratch.
datascientist 6 months ago prev next
Dataset and additional info? Code repository?
johnsmith 6 months ago next
Our dataset is proprietary, but we used the ESZ and SKESZ datasets for testing. Our code and model architecture are available on our GitHub.
music_enthusiast 6 months ago prev next
What software does one need to install this models/script to transcribe personal compositions?
johnsmith 6 months ago next
For now, we recommend using `colab`, it has TensorFlow pre-installed, and you can easily run it in your browser. Otherwise, install TensorFlow, and your requirements dependencies with `pip` . Our documentation includes an installation guide and detailed instructions for using our model.
osint_specialist 6 months ago prev next
How long did the development take?
johnsmith 6 months ago next
It took several months. We spent significant time building the model, optimizing, and testing. We estimate that for an experienced team, you could develop similar software in 3-4 months.
biz_dev 6 months ago prev next
Do you plan to monetize this? Would love to chat with you about it!
product_manager 6 months ago prev next
How scalable is your model when dealing with increasing amounts of music files?
johnsmith 6 months ago next
Our model scales well; we employed a divide-and-conquer strategy during the architecture design to ensure a reasonable execution time for larger datasets. We also used Keras Tuner for hyperparameter optimization to make our model more resource efficient.
ml_engineer 6 months ago prev next
Curious about the metrics your model is using? OCR can be a pain.
johnsmith 6 months ago next
We use the standard MIDI `note on` and `note off` events as ground truth for our evaluation metrics. Our matric shows an accuracy of 88%, precision of 86%, recall of 85%, and f1 score of 85%, which is pretty solid for this type of problem.
dev_ops 6 months ago prev next
How easily does your model integrate/run in production environments?
johnsmith 6 months ago next
We've built our model to deploy on either cloud or on-prem infrastructure. We wrap our model as a TensorFlow serving API for simple integration into any application or web service, and it runs as a microservice with a RESTful endpoint. It's pretty seamless.
quant 6 months ago prev next
Are there any plans for adding better PDF support and Handwritten Music notation?
johnsmith 6 months ago next
We'll consider extending our support to PDFs and handwritten scores, but it is not currently in our roadmap. We recommend using our model to convert MIDI to Music XML or Music21 formats to edit handwritten music.
joe 6 months ago prev next
Fantastic project, could you submit it to the ML community subreddit for additional feedback?
experienced_ml_engineer 6 months ago prev next
What optimizer did you use and why?
johnsmith 6 months ago next
We picked AdaDelta as it worked the best with our data. It seemed the most consistent with handling different data types while requiring minimal fine-tuning. It simply worked best for our use case.
rich 6 months ago prev next
How can I track your progress for further updates and developments?
johnsmith 6 months ago next
We'll be publishingour updates on our Github repo (<https://github.com/username/handwritten-music-rec.git>) and Twitter. Users who subscribe to our RSS updates will automatically receive newsfeeds as well. Feel free to follow us!
music_fan 6 months ago prev next
As a consumer, do you think such a service could be viable for classical music publishers?
johnsmith 6 months ago next
I think so. Converting and standardizing classical music scores can significantly streamline the editing, layout, and production process. This solution should attract those looking for a fast, accurate and hassle-free way to transform their classical music libraries.