TAFE-Net

This is the PyTorch Implementation for our paper:

TAFE-Net: Task-aware Feature Embeddings for Low Shot Learning (CVPR 2019)

Xin Wang, Fisher Yu, Ruth Wang, Trevor Darrell, Joseph E. Gonzalez

ArXiv link: https://arxiv.org/abs/1904.05967

Abstract

Learning good feature embeddings for images often requires substantial training data. As a consequence, in settings where training data is limited (e.g., few-shot and zero-shot learning), we are typically forced to use a generic feature embedding across various tasks. Ideally, we want to construct feature embeddings that are tuned for the given task. In this work, we propose Task-Aware Feature Embedding Networks (TAFE-Nets) to learn how to adapt the image representation to a new task in a meta learning fashion. Our network is composed of a meta learner and a prediction network. Based on a task input, the meta learner generates parameters for the feature layers in the prediction network so that the feature embedding can be accurately adjusted for that task. We show that TAFE-Net is highly effective in generalizing to new tasks or concepts and evaluate the TAFE-Net on a range of benchmarks in zero-shot and few-shot learning. Our model matches or exceeds the state-of-the-art on all tasks. In particular, our approach improves the prediction accuracy of unseen attribute-object pairs by 4 to 15 points on the challenging visual attribute-object composition task.

Requirements

• Pytorch (>0.4.0, tested on Pytorch 1.1.0)
• Scipy

Compositional Zero-shot Learning (attribute-object composition)

For the compositional zero-shot learning task, we are given a set of images and lists of attributes and objects. During training, only images of a subset of attribute-object composition are available. For testing, the model is tested on images of unseen (novel) attribute-object composition. We evaluate TAFE-Net on two datasets, MIT-States and StanfordVRD. For StanfordVRD, instead of attribute-object pairs, we instead use the subject-predicate-object(SPO) triplet as one composition.

Data Statistics

Download the processed features and labels from here and save the data to data/compositional-zs.

In the original paper, we evaluate the model with ResNet-101 as the feature extractor and in the recent paper Task-aware Deep Sampling, there are benchmark results of TAFE-Net with other backbones: ResNet-18, DLA-34 and DLA-102.

Dataset | Model | ResNet-18 @Top1 (%) | ResNet-101 @Top1 (%) |

To train the model, you can simply run

python3 -m compositional-zs.train_compose --cmd train --arch [ARCH] -cfg [ConfigName]

For testing,

python3 -m compositional-zs.train_compose --cmd test --resume [CheckpointPath]

To further simplify the procedure, you can write the configurations in configs.json. Some benchmark configs are already provided in configs.josn. You can just launch the job by specifying the configuration name as follows

python launch.py [ConfigName]

Zero-shot Learning

We evaluate TAFE-Net on five benchmark datasets: AWA1, AWA2, aPY, CUB and SUN. You can download the processed data here and save it to data/zero-shot. The processed data is originally from [3].

We provide benchmark results on both classic zero-shot learning (ZSL) and generalized zero-shot learning (GZSL) as follows.

For classic zer-shot learning (ZSL),

Models with * adopt synthetic features for training which could be considered as complementary approaches to our model as well as other discriminative models without * presented in the table.

For generalized zero-shot learning (GZSL),

Similar as the compositional zero-shot one, you can train the model by

python3 -m zero-shot.train_zsl --cmd train --arch [ARCH] -cfg [ConfigName]

and test by

python3 -m zero-shot.train_zsl --cmd test --resume [CheckpointPath]

Usually you can just launch the runs by specifying the config name in the config.json as

python launch.py [ConfigName]

Citation

If you find the code useful in your project, please consider citing our paper

@inproceedings{wang2019tafe,
  title={TAFE-Net: Task-Aware Feature Embeddings for Low Shot Learning},
  author={Wang, Xin and Yu, Fisher and Wang, Ruth and Darrell, Trevor and Gonzalez, Joseph E},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={1831--1840},
  year={2019}
}

Reference

[1] From Red Wine to Red Tomato: Composition with Context

[2] Attributes as Operators: Factorizing Unseen Attribute-Object Compositions

[3] Zero-Shot Learning - A Comprehensive Evaluation of the Good, the Bad and the Ugly

[4] Low-shot Visual Recognition by Shrinking and Hallucinating Features

[5] A Closer Look at Few-shot Classification