https://github.com/atrostan/4107_final_project
CNN model for classifying Psytrance Subgenres
https://github.com/atrostan/4107_final_project
cnn-keras genre-classification keras-tensorflow mir tensorflow-experiments
Last synced: about 2 months ago
JSON representation
CNN model for classifying Psytrance Subgenres
- Host: GitHub
- URL: https://github.com/atrostan/4107_final_project
- Owner: atrostan
- Created: 2018-12-16T21:34:30.000Z (over 6 years ago)
- Default Branch: master
- Last Pushed: 2018-12-17T02:19:48.000Z (over 6 years ago)
- Last Synced: 2025-03-16T16:51:28.237Z (2 months ago)
- Topics: cnn-keras, genre-classification, keras-tensorflow, mir, tensorflow-experiments
- Language: Jupyter Notebook
- Size: 21.3 MB
- Stars: 4
- Watchers: 0
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# 4107_Final_Project
Thanks to:
* https://gitlab.crowdai.org/gg12/WWWMusicalGenreRecognitionChallenge?_ga=2.123714251.1594486358.1545011930-1055586743.1544557330
* https://hackernoon.com/finding-the-genre-of-a-song-with-deep-learning-da8f59a61194
* https://github.com/Hvass-Labs/TensorFlow-Tutorials/blob/master/18_TFRecords_Dataset_API.ipynb
* https://ricardodeazambuja.com/deep_learning/2018/05/04/tuning_tensorflow_docker/
* https://medium.com/@moritzkrger/speeding-up-keras-with-tfrecord-datasets-5464f9836c36
* https://github.com/markjay4k/Fashion-MNIST-with-Keras/blob/master/pt3%20-%20FMINST%20Embeddings.ipynb
* https://github.com/keras-team/keras/blob/master/examples/conv_filter_visualization.py
# Requirements:
Seen in [`requirements.txt`](./py/requirements.txt)
|Library|Version|
|-------|-------|
|pandas|0.23.4|
|matplotlib|3.0.2|
|numpy|1.15.4|
|tensorflow|1.12.0|
|scipy|1.1.0|
|librosa|0.6.2|
|ipython|7.2.0|
|keras|2.2.4|
|scikit_learn|0.20.1|
# Download Tracks
Either download your own music from [http://www.ektoplazm.com/](http://www.ektoplazm.com/),
or
download music tracks from [`data_urls.txt`](./data_urls.txt)
# Convert Tracks to `TFRecords`
Detailed in [`read_mp3s.ipynb`](./read_mp3s.ipynb)
# Train model
Detailed in [`train_model.ipynb`](./train_model.ipynb).
You can skip this step, and just extract `best_model.tar.gz`: a Keras model
I've trained.
Load this model using [load_model(best_psy_cnn.h5)](https://stackoverflow.com/a/43263973/8822734)
# Test model
Detailed in [`test_model.ipynb`](./test_model.ipynb).
You can skip this step, and just extract `best_model.tar.gz`: a Keras model
I've trained.
# Predict using the model
Detailed in [`predict_model.ipynb`](./predict_model.ipynb).
The model was trained on 5 genres:
* [Downtempo](http://www.ektoplazm.com/style/downtempo)
* [Dark](http://www.ektoplazm.com/style/darkpsy)
* [Techno](http://www.ektoplazm.com/style/techno)
* [Goa](http://www.ektoplazm.com/style/goa)
* [Fullon](http://www.ektoplazm.com/style/full-on)
Download your own psytrance tracks (that can be characterized as one of the
above genres) and see how the model does!
To do this, make sure to compute the [mel-spectrogram](https://librosa.github.io/librosa/generated/librosa.feature.melspectrogram.html)
of your track, and feed the model with slices of your spectrogram.
# Visualize Convolutional Filters
Detailed in [`visualize_filters.ipynb`](./visualize_filters.ipynb) and
examples in [`./tex/diagrams/stitched_filters`](./tex/diagrams/)
# Report
The report outlining the development of this project is [here](./tex/report.pdf)
Production of graphs for the report are outlined in [`plot_results.ipynb`](./plot_results.ipynb)