NDDR-CNN, A Simple and General Architecture for General-Purpose CNN based Multi-Task Learning

---

Codes for ``NDDR-CNN: Layerwise Feature Fusing in Multi-Task CNNs by Neural Discriminative Dimensionality Reduction''.

NDDR-CNN is a simple and general architecture for general-purpose CNN based multi-task learning, which automatically learns feature fusing (i.e., sharing and splitting) at every CNN layer from different tasks. This is enabled by Neural Discriminative Dimensionality Reduction (NDDR), consisting of simple 1x1 convolution, Batch Normalization, and Weight Decay.

This repository illustrates example experiments (including ablations) with VGG-16 and VGG-16-Shortcut for semantic segmentation and surface normal prediction on NYUv2 dataset (reproduces Tables 1, 2, 3, 4, 6 of the paper).

Extension to other state-of-the-art CNN architectures and/or other tasks is similar and can be done straightforwardly in a “plug-and-play” manner.

Please refer to our paper for more technical details:

Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L. Yuille. NDDR-CNN: Layerwise Feature Fusing in Multi-Task CNNs by Neural Discriminative Dimensionality Reduction, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. [arXiv]

Usage

Prerequisites

• Python 2
• tensorflow (tested on 1.4.0 and 1.4.1), PIL, numpy, argparse. These can be installed by:

$ pip install pillow numpy argparse tensorflow-gpu==1.4.0

Dataset

First, prepare dataset and pretrained weights for the single task networks. Deeplab-VGG16 model and NYU v2 dataset are used as examples in this repository. You can download the official Caffe Deeplab-VGG16 model then converting it into tensorflow format with caffe-tensorflow, and the official NYU v2 dataset and the surface normal labels then formating them into specific separated folders with a list file indicating training/testing.

To make life easier, you can alternatively download the formatted datasets and the pretrained weights here, then unzip them in the current folder.

When you are all set, you should have the following sub-folds:

datasets/nyu_v2/list
datasets/nyu_v2/nyu_train_val
models/vgg_deeplab_lfov
models/nyu_v2/slim_finetune_seg
models/nyu_v2/slim_finetune_normal

Training

All the arguments to train/eval an NDDR-CNN are shown in util/input_arguments.py.

All the training scripts reproducing the results in Tables 1-4, 6 of our paper are in experiments folder, including ablation analysis and experiments on VGG-16, experiments on VGG-16-shortcut.

For example, you can simply use the following command to train an NDDR-CNN with VGG-16-shortcut architecture:

$ CUDA_VISIBLE_DEVICES=0 sh experiments/VGG16-shortcut/NDDR.sh

Evaluation

All the arguments to train/eval an NDDR-CNN are shown in util/input_arguments.py.

For example, to evaluate the model trained on VGG-16-shortcut architecture in the previous section, you can type:

$ CUDA_VISIBLE_DEVICES=0 python eval_multiple_tasks.py \
    --restore-from=save/VGG16_shortcut/NDDR \
    --task-1=seg --task-2=normal --num-classes-1=40 --num-classes-2=3 \
    --network=vgg_16_shortcut_deeplab_nddr

Contacts

For questions about the code or the paper, feel free to contact me by ethan.y.gao@gmail.com.

Bibtex

If this code is helpful to your research, please consider citing our paper by:

@inproceedings{gao2019nddr,
  title={NDDR-CNN: Layerwise Feature Fusing in Multi-Task CNNs by Neural Discriminative Dimensionality Reduction},
  author={Gao, Yuan and Ma, Jiayi and Zhao, Mingbo and Liu, Wei and Yuille, Alan L},
  booktitle = {CVPR},
  year={2019}
}

Acknowledgments

This project refers to the code from tf-slim models and tensorflow-deeplab-resnet.