Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

What makes ImageNet good for Transfer Learning?

This code provides a script to download our trained models for our paper 'What makes ImageNet good for Transfer Learning?' You can find our paper at or visit our project page


  • Python 2.7
  • nltk (optional)

Downloading trained models

To download all our models run the code.


The code will sequentially download all the models and save it the directory ./models

To specify which models to download you can run the code with the flag -e

# Downloads all the models for the hierarchy experiment
python -e hierarchy

# Downloads all the models from the class experiment
python -e class

To go one step further and select a specific experiment you can pass the flag -s

# Downloads the model from the hierarchy experiment trained with 918 classes
python -e hierarchy -s 918

You can also specify the save destination using the flag -d

# Save the specific model on the destination ./dst
python -e hierarchy -s 918 -d ./dst

Label set generator and WordNet tools

We provide the code to generate the label sets mentioned in the hierarchy experiments.

To generate the label sets yourself, you will need to download the python compatible WordNet corpus from
For some python examples on how to use WordNet, refer to

We also include helper functions to easily manipulate wordnet tree in

To run the script to generate label sets


You can also specify the experiment

# By default -e is set to all
python -e up_down
python -e down_up
python -e all 

PASCAL / ImageNet common classes

Common classes shared between PASCAL and ImageNet can be found here

LMDB generation toolkit (coming soon)

We will soon provide tools to generate LMDB file for Caffe

Experiment Setup (coming soon)

We will soon release helper code for setting up experiments for arbitrary networks.
ResNet-34 models for the experiments above will be released soon.


models and tools for -What makes ImageNet good for Transfer Learning?






No releases published


No packages published