Warning

This repository is archived and is no longer maintained.

Conveiro

Conveiro (convolutional + oneiro, Greek for "dream") is an open source library for feature visualization in deep convolutional networks. It implements multiple techniques for visualization, such as laplace, multiscale, deep dream and CDFS.

All of these methods are based on:

Deep dream

Deep dream is implementation of technique based on

• https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/tutorials/deepdream/deepdream.ipynb
• https://ai.googleblog.com/2015/06/inceptionism-going-deeper-into-neural.html

How it works:

• We create random image (or we can use seed image)
• We feed this image to network and optimize it based on calculated gradients
• We employ few clever tricks based on scaling and frequencies

There are few more steps but this is the essence of this technique.

CDFS

CDFS (color-decorrelated fourier space) is custom implementation of technique based on

• https://distill.pub/2017/feature-visualization/
• https://github.com/tensorflow/lucid

How it works:

• We generate random complex coefficient
• We use said coefficients to generate image by inverse fourier transformation
• After we feed this image to network we can calculate gradients and use gradient descent to optimize these coefficient

There are few more steps but this is the essence of this technique.

Requirements

• Python 3.4 and above
• Tensorflow (CPU or GPU variant, version 2 not yet supported)
• Numpy
• Matplotlib
• click, tensornets, pillow, graphviz (if you want to use the command-line tool with examples)

Installation

pip install conveiro

Development version

pip install -e .    # from cloned repository

Command-line usage

This library comes with a command-line tool called conveiro that can visualize and hallucinate networks from tensornets library.

Usage: conveiro COMMAND [OPTIONS] [ARGS]...

Commands:
  graph     Create a graph of the network architecture.
  layers    List available layers (operations) in a network.
  networks  List available network architectures (from tensornets).
  render    Hallucinate an image for a layer / neuron.

Run conveiro --help or conveiro [command-name] --help to show the list of capabilities and options.

Examples

For examples how to use this library please take a look at jupyter notebooks in docs/ folder:

• https://github.com/Showmax/conveiro/tree/master/docs/deep_dream.ipynb
• https://github.com/Showmax/conveiro/tree/master/docs/cdfs.ipynb

Simplest example:

import tensorflow as tf
import tensornets as nets
from conveiro import cdfs

input_t, decorrelated_image_t, coeffs_t = cdfs.setup(224)

model = nets.Inception1(input_t)
graph = tf.get_default_graph()

with tf.Session() as sess:
    sess.run(model.pretrained())

    objective = graph.get_tensor_by_name("inception1/block3b/concat:0")
    image = cdfs.render_image(sess, decorrelated_image_t, coeffs_t, objective[..., 55], 0.01)
    cdfs.show_image(cdfs.process_image(image))

Note The API is preliminary and may change in future versions.

Articles

• Showmax Tech Blog: Introducing Conveiro...