A flexible variational inference LDA library.
Switch branches/tags
Nothing to show
Clone or download

README.md

LDA++

LDA++ is a C++ library and a set of accompanying console applications that enable the inference of various Latent Dirichlet Allocation models.

The project provides three console applications

  • lda implementing LDA
  • slda implementing Supervised LDA
  • fslda implementing Fast Supervised LDA

and a library that can be used from your own C++ projects.

You can read the documentation site at ldaplusplus.com and there is of course an API documentation as well.

How to get it

We use CMake for building the project and currently only provide the option to build from source. The LDA++ installation process is straightforward and documented at our site.

Console applications

We expect that the preferred way of using LDA++ will be through the provided console applications. You can read thorough documentation for them as well. All our console applications are designed to read matrix files serialized in numpy format so that one can easily create files in a python session.

It suffices to say that the following shell session runs the Fast Supervised LDA (fsLDA) on the scikit learn digits dataset (provided you have installed LDA++).

$ python
Python 2.7.12 (default, Jul  1 2016, 15:12:24) 
[GCC 5.4.0 20160609] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> from sklearn.datasets import load_digits
>>> import numpy as np
>>> d = load_digits()
>>> with open("digits.npy", "wb") as f:
...     np.save(f, d.data.astype(np.int32).T)
...     np.save(f, d.target.astype(np.int32))
... 
>>> exit()
$ fslda train digits.npy model.npy
E-M Iteration 1
100
200
300
400
500
600
700
800
900
1000
1100
1200
1300
1400
1500
1600
1700
log p(y | \bar{z}, eta): -4137.75
log p(y | \bar{z}, eta): -3230.67
log p(y | \bar{z}, eta): -2758.81
log p(y | \bar{z}, eta): -2498.32
log p(y | \bar{z}, eta): -2341.4
log p(y | \bar{z}, eta): -2240.48
log p(y | \bar{z}, eta): -2172.38
log p(y | \bar{z}, eta): -2124.71
log p(y | \bar{z}, eta): -2090.4
log p(y | \bar{z}, eta): -2065.15
...
$ python
Python 2.7.12 (default, Jul  1 2016, 15:12:24) 
[GCC 5.4.0 20160609] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy as np
>>> with open("model.npy") as f:
...     alpha = np.load(f)
...     beta = np.load(f)
...     eta = np.load(f)
...
>>> import matplotlib.pyplot as plt
>>> plt.imshow(beta[0].reshape(8, 8), interpolation='nearest', cmap='gray')
<matplotlib.image.AxesImage object at 0x7f4cf201b810>
>>> plt.show()
>>> exit()

License

MIT license found in the LICENSE file.