Measuring Statistics about Information Sources in MIDI Files

https://nbviewer.jupyter.org/github/craffel/midi-ground-truth/blob/master/Statistics.ipynb

Measuring the Reliability of MIDI-Derived Annotations

https://nbviewer.jupyter.org/github/craffel/midi-ground-truth/blob/master/Reliability%20Experiments.ipynb

Optimizing DTW-Based Audio-to-MIDI Alignment and Matching

https://nbviewer.jupyter.org/github/craffel/alignment-search/blob/master/overview.ipynb

Lakh MIDI Dataset Tutorial

https://nbviewer.jupyter.org/github/craffel/midi-dataset/blob/master/Tutorial.ipynb

Misc

http://millionsongdataset.com/ https://docs.google.com/document/d/1mZDY0XUZGtemI6zgCSJzQ8MzL1MqPX8WNTkuoNxBd5g/edit https://towardsdatascience.com/bachgan-using-gans-to-generate-original-baroque-music-10c521d39e52

https://www.intel.com/content/www/us/en/artificial-intelligence/posts/deep-learning-training-and-inference.html

https://console.firebase.google.com/u/1/project/music-generator-9578a/overview

https://www.anaconda.com/

Repo's

https://github.com/salu133445/musegan

https://salu133445.github.io/pypianoroll https://salu133445.github.io/muspy

https://github.com/salu133445/ismir2019tutorial

Libs

https://docs.aiohttp.org/en/stable/index.html https://fastapi.tiangolo.com/

SCREEN CHEAT SHEET - https://kapeli.com/cheat_sheets/screen.docset/Contents/Resources/Documents/index

REST API

https://github.com/pytorch/tutorials/blob/master/intermediate_source/flask_rest_api_tutorial.py https://pytorch.org/tutorials/intermediate/flask_rest_api_tutorial.html https://libiseller.ai/deploying-pytorch-to-production

Notes

• Don't use pipenv (https://chriswarrick.com/blog/2018/07/17/pipenv-promises-a-lot-delivers-very-little/)
  • Here are some alternatives
    • https://virtualenvwrapper.readthedocs.io/en/latest/
    • https://github.com/jazzband/pip-tools
    • https://github.com/python-poetry/poetry

Other intresting links

Fake anime using DCGAN - https://www.youtube.com/watch?v=cqXKTC4IP10 Building our first simple GAN - https://www.youtube.com/watch?v=OljTVUVzPpM Sentdex - https://www.youtube.com/channel/UCfzlCWGWYyIQ0aLC5w48gBQ/videos

https://github.com/onnx/tutorials

https://discuss.pytorch.org/t/runtimeerror-cudnn-error-cudnn-status-not-initialized/115286/2

https://github.com/MicrosoftDocs/azure-docs/blob/master/articles/machine-learning/how-to-train-pytorch.md

This version of torch

(1.7.0+cu101 - https://download.pytorch.org/whl/torch_stable.html)

Different states of this application

This project is used to train our model, run predictions on our model and handle restful API requests (responding with predictions). I need to find a way to easily:

• Train
• Predict without API
• Start web service with access to predictions

Prerequisites

• Make sure that you have python installed
• Make sure that you have the packages pip, virtualenv installed
• Install Fluidsynth
• Data
  • Download LPD dataset if creating new tensor dataset
  • Download .pt if loading existing tensor dataset for training model
  • Download checkpoints loading exisiting model
  • Download synth files for midi2audo (Optional, this uses Fluidsynth)

Installation Steps

1. Clone this repo & cd into it
2. Create a virtualenv & activate it
3. Install pip-tools
4. Run pip-sync to install packages
   • This torch version is installing with cuda, if you don't have a GPU which is supported by cuda than install the CPU version of torch.
5. Run python main.py

TODO

• Make sure all outputed files are uniquely named
• Make sure these functions work

  • Gluing and trimming midi for a certain predetermined time in seconds, implemented the ability to use templates in the file name, so only generated midis can be glued in one of 4 ways

  • Save/download end-points for both .mid and .wav

  • ai3 & ai4 are not ai but they are "reverse playback" & "tonal inversion" (Don't know which order).

    • Message from Upwork "This is not ai in fact, just reverse playback and tonal inversion in the MuseGAN_torch_main file these functions"
    • But if this is only reverse playback or tonal inversion than the bellow sentance in the tech docs he sent me don't make sense
      • "The genre field allows you to select the genre of generated music, (this option is only available when working with ai3 and ai4"

  • Maybe the service which is responisble for generating a file name should be some service, which takes care of loading, saving, naming and etc safeley and can be used [DONE]

  • Service for orchestrating the neccesary actions[DONE]

  • Replacing notes with chords, you can set in which track or tracks to change and select a major or minor triad [DONE]

  • Changing instruments by tracks [DONE]

  • Convert midi to wav using various sound fonts in sf2 format. [DONE]

Notes

• Each generation request is unique and delivered songs can't be modified but they can be downloaded again.

• The genre portion of the bellow generation proccess is removed until I figure it out but here are soem notes

  • Genre will most likley dictate which pre-trained model (Different trained models for different genre's) is used so it would need to one of the first operations

• We will mainly focus on returning the WAV but if the user wants to save the midi as well they can request it with the file name

• Generation proccess

  • path, filename = getFilename()

  • output_npz_file = predict(generator, filename)

  • current_file = output_npx_file

  • requested_operations = { change_instruments: { 1: 0, 2: 0, 3: 27 }, add_drums: True, add_chords: True, set_bpm: 100, modify_length: 260, tone_invert: True, invert_midi: True, }

  • available_operations = { change_instruments: change_instruments, add_drums: add_drums, add_chords: add_chords, set_bpm: set_bpm, modify_length: modify_length, tone_invert: tone_invert, invert_midi: invert_midi, }

  • for operation in available_operations: if operation in requested_operations: operation_value = requested_operations[operation] current_file = operation(operation_value)

  • output_wav_file = convert_midi2_wav(current_file)

  • return output_wav_file

• WRITE OPERATIONS TO HAPPEN WHEN THEY NEED TO IN A LOOP SO YOU ONLY NEED TO LOOP ONCE AND SAVE ONCE

• (Do this in your own project)

Firebase

• How to auth (https://firebase.google.com/docs/auth)

  • This is more of an overview of how to implement firebase auth into your client (ios, android, web (front-end) and etc...)
  • Quotes from above

    • To sign a user into your app, you first get authentication credentials from the user.

      • These credentials can be the user's email address and password, or an OAuth token from a federated identity provider.

    • Then, you pass these credentials to the Firebase Authentication SDK.

      • Our backend services will then verify those credentials and return a response to the client.

    • After a successful sign in, you can access the user's basic profile information

    • You can also use the provided authentication token to verify the identity of users in your own backend services.

• Verify id tokens (https://firebase.google.com/docs/auth/admin/verify-id-tokens)

  • This is what I'll be doing in the back-end to create and manage users and etc

  • Take the uid and try to get the user, if it dosen't exist create it.

  • Than do whatever you need to do with the uid / user.