This fork added additional classes to analyze spectrograms of classical piano works. The audio files were taken from the Classical Piano Midi Page and processed by the utility classes in the utilities sub-project. The com.github.tjake.rbm.music package contains classes based on their com.github.tjake.rbm.minst counterparts, but they should also be generic enough that they can also analyze the MNIST data set with little modification.
Other changes include making some of the classes in com.github.tjake.rbm a little more generic, as well as general formatting changes.
- MP3 files were downloaded from the Classical Piano Midi Page
- MP3 audio file were converted to WAV via the
ConvertAllMp3ToWavclass - WAV files were imported into Virtual ANS and exported as PNG spectrograms (for audio, Virtual ANS only reads the WAV file type)
- Spectrograms were cropped, scaled down, and partitioned by the
ScaleAndPartitionSpectrogramclass - those new PNG files were added to the
resources/testandresources/traindirectories
Learn more about this project from this blog post:
http://tjake.github.com/blog/2013/02/18/resurgence-in-artificial-intelligence/
This project provides a implementation for a Restricted Boltzmann Machine and a Deep Belief Network
It uses the MNIST handwritten dataset to illistrate an example RBM and DBN.
From source build the project with maven:
- mvn deploy
This will build a single jar and download the mnist dataset.
- java -jar target/rbm-dbn-mnist-0.0.1.jar
Runs the app. shows the usage screen
Usage: [rbm minst-labels.gz minst-images.gz]
[dbn minst-images.gz minst-labels.gz dbn.bin]
[gen dbn.bin]
- java -jar target/rbm-dbn-mnist-0.0.1.jar rbm target/minst/train-labels-idx1-ubyte.gz target/minst/train-images-idx3-ubyte.gz
Trains a single RBM with 100 hidden nodes. Each of the hidden nodes weights are rendered alongside the test digit in blue.
- java -jar target/rbm-dbn-mnist-0.0.1.jar dbn target/minst/train-labels-idx1-ubyte.gz target/minst/train-images-idx3-ubyte.gz /tmp/dbn.bin
Trains a Deep Belief Network made up of three RBMs. It learns to match pictures of digits with their corresponding label. It takes about 10m to train but once it's done it has ~95% accuracy rate. The trained DBN is saved to a file.
- java -jar target/rbm-dbn-mnist-0.0.1.jar gen /tmp/dbn.bin
Takes the trained DBN from step 4. and reverses the flow, generating a visual image of a digit from a digit label.
Copyright 2013 T Jake Luciani [email protected]
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the 'Software'), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED 'AS IS', WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.