Code for Net2Vec: Quantifying and Explaining how Concepts are Encoded by Filters in Deep Neural Networks
Switch branches/tags
Nothing to show
Clone or download
Pull request Compare This branch is 102 commits ahead, 4 commits behind CSAILVision:release1.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
.gitignore working on README Feb 8, 2018



This repository contains the code for our arxiv'18 paper Net2Vec: Quantifying and Explaining how Concepts are Encoded by Filters in Deep Neural Networks. It is forked code-wise and builds on the work from Bau et al's CVPR'17 paper Network Dissection: Quantifying Interpretability of Deep Visual Representations.

Pardon the current appearance of the repo: this code is still being developed and will be cleaned up (with more user-friendly README instructions) shortly.

Probing a Caffe/Pytorch Network

First, collect network activations using either src/ to prove a Caffe network or src/ to probe a PyTorch network. For PyTorch networks, blobs are based on the path to the module of interest, with "." denoting entering into a nn.Module/nn.Sequential (i.e., to probe conv5 in the pytorch implementation of pytorch, pass in "features.11" as the blob name).

Second, collect activation quantiles as done in NetDissect using src/



Run src/ to probe network activations using single filters (use this regardless of if the network was from Caffe or PyTorch; this function approximately does the same thing as the original src/ but we save results in a different format and have a few minor implementation differences (i.e., 1., upsampling after thresholding for consistently with the multi-filter approach, 2., upsampling bilinearly without respect to the receptive field, 3., using the BRODEN train/val split to choose the best filter, see our paper for more details).


Run src/ first and then src/

To be continued...


To be continued...

README from NetDissect

You can use this code with naive Caffe, with matcaffe and pycaffe compiled. We also provide a PyTorch wrapper to apply NetDissect to probe networks in PyTorch format. There are dissection results for several networks at the project page.

This code includes

  • Code to run network dissection on an arbitrary deep convolutional neural network provided as a Caffe deploy.prototxt and .caffemodel. The script runs all the needed phases.

  • Code to create the merged Broden dataset from constituent datasets ADE, PASCAL, PASCAL Parts, PASCAL Context, OpenSurfaces, and DTD. The script runs all the needed steps.


  • Clone the code of Network Dissection from github
    cd NetDissect
  • Download the Broden dataset (~1GB space) and the example pretrained models.

Note that you can run script/ to download Broden dataset with images in all three resolution (227x227,224x224,384x384), or run script/ to download more CNN models. AlexNet models work with 227x227 image input, while VGG, ResNet, GoogLeNet works with 224x224 image input.

Run in Caffe

  • Run Network Dissection in Caffe to probe the conv5 layer of the AlexNet trained on Places365. Results will be saved to dissection/caffe_reference_model_places365/, in which html contains the visualization of all the units in a html page and conv5-result.csv contains the raw predicted labels for each unit. The code takes about 40 mintues to run, and it will generate about 1.5GB intermediate results (mmap) for one layer, which you could delete after the code finishes running.
    script/ --model caffe_reference_places365 --layers "conv5" --dataset dataset/broden1_227 --resolution 227
  • Run Network Dissection to compare three layers of AlexNet trained on ImageNet. Results will be saved to dissection/caffe_reference_model_imagenet/.
    script/ --model caffe_reference_imagenet --layers "conv3 conv4 conv5" --dataset dataset/broden1_227 --resolution 227
  • If you need to regenerate the Broden dataset from scratch, you can run script/ The script will download the pieces and merge them.

  • Network dissection depends on scipy as well as pycaffe. Details on installing pycaffe can be found here.

Run in PyTorch

  • Or try script/ on a resnet18 trained on Places365.


  • At the end of the dissection script, a report will be generated that summarizes the semantics of the networks. For example, after you have tested the conv5 layer of caffe_reference_places365, you will have:

These are, respectively, the HTML-formatted report, the semantics of the units of the layer summarized as a bar graph, visualizations of all the units of the layer (using zero-indexed unit numbers), and a CSV file containing raw scores of the top matching semantic concepts in each category for each unit of the layer.


If you find the code useful, please cite the following papers

  title={Net2Vec: Quantifying and Explaining how Concepts are Encoded by Filters in Deep Neural Networks},
  author={Fong, Ruth and Vedaldi, Andrea},
  journal={arXiv preprint arXiv:1801.03454},

  title={Network Dissection: Quantifying Interpretability of Deep Visual Representations},
  author={Bau, David and Zhou, Bolei and Khosla, Aditya and Oliva, Aude and Torralba, Antonio},
  booktitle={Computer Vision and Pattern Recognition},