This folder contains the code used in the paper. The version of TensorFlow (0.12) is now very outdated, but we are still including the original code here for reference purposes.
The code is structured as follows
experiments/ # Example configs (actual configs generated in tools/scripts/)
transfers/ # Example transfer config
aws_second/ # Task-specific configs used for paper
final/ # Configs used for taskbank (slightly improved)
lib/ # The bulk of the TF code
data/ # Dataloading code
losses/ # Different types of losses (e.g. for GANs)
models/ # Architectures
optimizers/ # Ops and train_steps. E.g. for GANs
savers/ # Code for saving checkpoints to S3
notebooks/ # Jupyter notebooks used for developing
*.ipynb # Debugging notebooks, hyperparameter sweeps, etc
analysis/ # Notebooks that analyze aggregated transfers. Early implementation of BIP solver
transfer_viz/ # Visualizing results of networks with different HPs
quality_control/ # Visualizing task-specific network outputs
bottleneck/ # Results with different bottleneck sizes
quality_control_final # Visualizing taskbank networks
tools/ # Utilities
scripts/ # Scripts to generate configs
extract_losses.py # Compute losses on a train/val/test set
train.py # Training script
transfer.py # Transfer script- Create all task-specific configs (using
tools/scripts/and saving toexperiments/) - Select which config to run, then train the task-specific network using
tools/train.py - Create all transfer configs (using
tools/scripts/and saving toexperiments/) - Select which config to run, then train the transfer network using
tools/transfer.py - Compute losses with
extract_losses.py - Generate win rates and affinities with one of the methods in
analysis/
Note: this folder provides the full code and additional resources for archival and information purposes only. We dont maintain the code here. For trained TASK BANK networks and demo code for running them, please see the TASK BANK folder. For Taskonomy dataset, please see the DATASET folder. For more details and the full methodology, please see the main paper and website.
If you find the code, models, or data useful, please cite this paper:
@inproceedings{zamir2018taskonomy,
title={Taskonomy: Disentangling Task Transfer Learning},
author={Zamir, Amir R and Sax, Alexander and and Shen, William B and Guibas, Leonidas and Malik, Jitendra and Savarese, Silvio},
booktitle={2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2018},
organization={IEEE}
}
