DarkSight is a dimension reduction technique to visualize any black-box classifier.

Please visit here for more information (the related paper, main results, exemplar visualization and demos).

This repository contains a PyTorch implementation of DarkSight.

Download or clone this repository and put the folder darksight in the same path as the Python file need to import it.

• Python 2.7.12
  • Python 3.x is not supported
• PyTorch 0.3.1
  • PyTorch 0.2.x is not supported
• NumPy 1.13.3
• Matplotlib 2.0.2

A demo is provided in this repository to illustrate the basic use of DarkSight, using the provided output from a 98.2% accuracy LeNet trained on MNIST dataset.

darksight.Knowledge(output, T=1) is the class that wraps a clssifier's output.

Parameters

• output: the classifier's output; should be a numpy.array
  • It can be either the predictive probabilities or the logit before softmax

Optional

• T: the temperature used when normalizing the provided output

darksight.DarkSight(klg) is the class for the DarkSight proposed in the paper. It defines a Naive Bayes model on 2D and performs non-parametric dimension reduction and model compression jointly based on a symmetric KL divergence objective.

Parameters

• klg is the knowledge object wrapped by darksight.Knowledge

Parameters

• num_epoch: number of epochs for training
• lrs: learning rates for each component, conditional distribution, low-dimensional embedding and prior distribution, in a list, i.e. [lr_cond, lr_y, lr_prior]

Optional

• batch_size: batch size for traning
• verbose_skip: number of epochs for printing training logs
• do_annealing: whether to do annealing or not
• annealing_length: length of epoch for annealing
• highest_T: start temperature of annealing
• annearling_stepsize: the step size of calculating new temperature

Helper function to plot the loss trace.

Helper function to print the learnt low-dimensional embeddings as a scatter plot.

Optional

• color_on: color or monotonic?
• mu_on: plot means of each cluster?
• labels: a Python list of string for labels used in the plot
• contour_on: plot contour based on P(y)?
  • Note that generating contour requires running the model so we need to specify whether and how to use GPU for this purpose below
• use_cuda: use GPU?
• gpu_id: which GPU to use?
• contour_slices: how fine do you want your contour to be computed?
• contour_num: how many contour levels you want?

Low-dimensional embedding learnt.

Means of each conditional distribution learnt.

Precision matrix of each conditional distribution learnt.

Parameters of the prior distribution.

Helper function to output the training results.

Parameters

• output_file_path: output file path

Kai Xu is the first author of the corresponding paper and the maintainer of this library, feel free to contact him on the paper or the library by email: [email protected].