keravis 1.0.2

A high-level API for ConvNet visualizations in Keras

keravis

keravis is a high-level API for ConvNet visualizations in Keras. As of v1.0, it supports visualizations of

1. Convolutional layer activations
2. 2-dimensional feature space representations
3. Saliency maps (vanilla backprop, guided backprop, and occlusion)
4. Generated inputs that result in maximal class scores
5. Patches in a set of images that maximally activate an intermediate neuron

with support for nested pretrained models.

This is a hobby project that was inspired by lecture 14 of Stanford's CS231n: Convolutional Neural Networks for Visual Recognition http://cs231n.stanford.edu/. It is not yet optimized for serious use (see keras-vis instead).

Installation

You can install keravis using pip

pip install keravis

Usage

Read the documentation

Sample Visualizations

Below are sample visualizations from a small convolutional network trained on MNIST

from keravis import feature_space
feature_space(model,X=x_test[:5000],y=y_test[:5000],kind='tsne')

from keravis import saliency_backprop
saliency_backprop(model,test_img,class_idx=7)

from keravis import saliency_guided_backprop
saliency_guided_backprop(model,test_img,class_idx=7)

from keravis import maximal_class_score_input
maximal_class_score_input(model,class_idx=5,dim=(28,28,1))

from keravis import maximally_activating_patches
maximally_activating_patches(model,'conv2d_1',X=x_test)