latentplot 0.0.2

Description

Python package to plot the latent space of a set of images with different methods.

Install with pip

$ python3 -m pip install latentplot --user

Install from source

$ git clone https://github.com/luiscarlosgph/latentplot.git
$ cd latentplot
$ python3 setup.py install --user

Exemplary code snippet

# List of BGR images of shape (H, W, 3)
images = [ ... ]           

# List of vectors of shape (D,), where D is the vector dimension
feature_vectors = [ ... ]  

# List of integer class labels
labels = [ ... ]           

# Produce a BGR image containing a 2D plot of the latent space with t-SNE
plotter = latentplot.Plotter(method='tsne')  # You can use either 'pca', 'tsne' or 'umap'                              
im_tsne = plotter.plot(images, feature_vectors, labels)  # Providing labels is optional

The latentplot.Plotter constructor parameters are:

Exemplary results

• CIFAR-10: the size of the images in this dataset is 32x32 pixels. The colour of the rectangle around each image indicates the class label of the image. The colour for each class is randomly chosen in every run.

  • PCA:

    

  • t-SNE:

    

  • UMAP:

    

Notes on dimensionality reduction methods

• PCA (principal component analysis):

  Assumes correlation (linear relationship) between features, sensitive to the scale of the features (features whose range is wider are more likely to become principle components), and it is not robust to outliers.

• t-SNE (t-Distributed Stochastic Neighbor Embedding):

  t-SNE does not assume linear relationships between features. Observations that are close in the high-dimensional space are expected to be close in the dimensionality-reduced space. t-SNE copes well with odd outliers.

• UMAP (Uniform Manifold Approximation and Projection):

  Fast, and scales well with regards to both dataset size and dimensionality.

Author

Luis Carlos Garcia Peraza Herrera (luiscarlos.gph@gmail.com), 2023.

License

This code repository is shared under an MIT license.