Acai is an open source project initialised by Berry Labs, a startup working towards machine learning algorithms to solve daily issues. Acai (codename) is trying to solve the problem of The Tyranny of Choice (a.k.a Paradox of Choice) to describe the misery of users facing over-abundant choices. In the music area, especially in the age of streaming music, this paradox becomes so significant that it affects every single piece of choice when users try to enjoy music. It's why this project was born. http://www.acai.berry.ai/
These instructions will get you a copy of the project up and running on your local machine for development and testing purposes. See deployment for notes on how to deploy the project on a live system.
What things you need to install the software and how to install them
Python
Python-Librosa
Python-Lutorpy
Python-NumPy
Python-Scipy
Torch
Torch-cunn
Torch-dp
Torch-nn
Torch-optim
Torch-xlua
All major distributions of Linux provide packages for both Python and NumPy.
pip install librosa
pip install lutorpy
pip install numpy
pip install scipy
Personally, I will recommend Anaconda as default Python compiler. To install them, go to page
https://www.continuum.io/downloads
and find the proper install packages
You can find torch installation instruction in the official site: http://torch.ch/ You can also refer to our wiki page: https://github.com/BerryAI/music_cortex/wiki/Torch-Setup
In this project, we use some public open database, and they are
- Last.fm 1k user data download
- Million Song Databasedownload
- Million Song Database subset download
- Echo Nest user data download
For convenience purpose, I have calculate the intersection between 1k user data and MSD Database. HERE is the download link.
Test functions are under ./test folder. After downloading all the data files, please put the extracted files into ./data folder.
Then run collaborative filtering
python test_cf_hf_gd.py
run convolutional neural networks
th example.lua
in command line under the directory of the project installed.
The recommendation equation is:
Where U is the full set of all users, 
- K nearest neighbours:
Where
is the set of neighbors of user a.
- Pearson Correlation:
* Matrix Factorization and Hidden Features
We could use much smaller dimension matrix P, Q to represent and approximate the
full rating score matrix R. That is:
Normally, we have two different approaches:
- Singular Value Decomposition
R is a m*n matrix
Where:
- M is m*m unitary matrix
- Σ is m*n diagonal matrix with singular values
- N is n*n unitary matrix
With first k singular values, we could approximate R as:
Then:
- Gradient Descent
We try to minimize the norm of residue matrix:
we have two different approaches:
-
Classic Gradient Descent:
-
Stochastic Gradient Descent with momentum:
Both methods will converge, but please be careful choosing coefficients.
1.Convolutional Layer
Convolutional layers are the core building block of CNNs. The layer's parameters consist of sets of learnable filters/kernels 
2.Max Pooling Layer
Pooling layer is another important concept of CNN. It is down-sampling process. Max pooling is a non-linear down-sampling method. The forward equation is:
3.Rectified Linear Units Layer
ReLU layers apply nonlinear activation function to neurons. Comparing to other common activation functions, ReLU is fast in training and suffers less on gradient extenuation during training.The forward equation is:
4.Loss Layer The loss layer is the last layer in CNN which defines the training deviation between real predicted results and target results. We provide 2 options in our model.
-
Mean Squared Error
where
is real output and is target output.
-
Softmax Loss
where
- Stochastic Gradient Descent with momentum:
- Fork it!
- Create your feature branch:
git checkout -b my-new-feature - Commit your changes:
git commit -am 'Add some feature' - Push to the branch:
git push origin my-new-feature - Submit a pull request :D
The OpenMRS source code and binaries are released under the MIT license