Label Hallucination for Few-Shot Classification

Last update: Nov 13, 2022

Related tags

Overview

Label Hallucination for Few-Shot Classification

This repo covers the implementation of the following paper: Label Hallucination for Few-Shot Classification . If you find this repo useful for your research, please consider citing the paper.

@article{Jian2022LabelHalluc,
    author = {Yiren Jian and Lorenzo Torresani},
    title = {Label Hallucination for Few-shot Classification},
    journal = {AAAI},
    year = {2022}
}

@article{jian2021label,
      title={Label Hallucination for Few-Shot Classification},
      author={Yiren Jian and Lorenzo Torresani},
      journal={arXiv preprint arXiv:2112.03340},
      year={2021}
}

Requirements

This repo was tested with Ubuntu 18.04.5 LTS, Python 3.6, PyTorch 1.4.0, and CUDA 10.1. You will need at least 32GB RAM and 22GB VRAM (i.e. two Nvidia RTX-2080Ti) for running full experiments in this repo.

Download Data

The data we used here is preprocessed by the repo of MetaOptNet, Please find the renamed versions of the files in below link by RFS.

Download and unzip the dataset, put them under data directory.

Embedding Learning

Please follow RFS, SKD and Rizve et al. (or other transfer learning methods) for the embedding learning. RFS provides a Dropbox link for downloading their pre-trained models for miniImageNet.

We provide our pretrained embedding models by [SKD] and [Rizve et al.] at Dropbox. Note that those models are NOT the official release by original authors, and they perform slightly worse than what reported in their papers. Better models could be trained with longer durations and/or by hyper-parameters tuning.

Once finish the embedding training, put the pre-trained models in models_pretrained directory.

Running Our Fine-tuning

To perform 5-way 5-shot classifications, run:

# For CIFAR-FS
CUDA_VISIBLE_DEVICES=0 python -W ignore eval_fewshot_SoftPseudoLabel.py --dataset CIFAR-FS --data_root data/CIFAR-FS/ --model_path models_pretrained/cifar-fs_skd_gen1.pth --n_shot 5 --n_aug_support 5 --epoch 1 --norm_feat

# For FC100
CUDA_VISIBLE_DEVICES=0 python -W ignore eval_fewshot_SoftPseudoLabel.py --dataset FC100 --data_root data/FC100/ --model_path models_pretrained/fc100_skd_gen1.pth --n_shot 5 --n_aug_support 5 --epoch 1 --norm_feat

# For miniImageNet (require multiple GPUs)
CUDA_VISIBLE_DEVICES=0,1 python -W ignore eval_fewshot_SoftPseudoLabel.py --dataset miniImageNet --data_root data/miniImageNet/ --model_path models_pretrained/mini_skd_gen1.pth --n_shot 5 --n_aug_support 5 --epoch 1 --norm_feat

# For tieredImageNet (require multiple GPUs)
CUDA_VISIBLE_DEVICES=0,1 python -W ignore eval_fewshot_SoftPseudoLabel_tieredImageNet.py --dataset tieredImageNet --data_root data/tieredImageNet/ --model_path models_pretrained/tiered_skd_gen0.pth --n_shot 5 --n_aug_support 5  --early 200 --print 50 --norm_feat

To perform 5-way 1-shot classifications, run:

# For CIFAR-FS
CUDA_VISIBLE_DEVICES=0 python -W ignore eval_fewshot_SoftPseudoLabel.py --dataset CIFAR-FS --data_root data/CIFAR-FS/ --model_path models_pretrained/cifar-fs_skd_gen1.pth --n_shot 1 --n_aug_support 25 --epoch 3 --norm_feat

# For FC100
CUDA_VISIBLE_DEVICES=0 python -W ignore eval_fewshot_SoftPseudoLabel.py --dataset FC100 --data_root data/FC100/ --model_path models_pretrained/fc100_skd_gen1.pth --n_shot 1 --n_aug_support 25 --epoch 5 --norm_feat

# For miniImageNet (require multiple GPUs)
CUDA_VISIBLE_DEVICES=0,1 python -W ignore eval_fewshot_SoftPseudoLabel.py --dataset miniImageNet --data_root data/miniImageNet/ --model_path models_pretrained/mini_skd_gen1.pth --n_shot 1 --n_aug_support 25 --early 150 --norm_feat

# For tieredImageNet (require multiple GPUs)
CUDA_VISIBLE_DEVICES=0,1 python -W ignore eval_fewshot_SoftPseudoLabel_tieredImageNet.py --dataset tieredImageNet --data_root data/tieredImageNet/ --model_path models_pretrained/tiered_skd_gen0.pth --n_shot 1 --n_aug_support 25  --early 200 --print 50 --norm_feat

Reading the outputs

400it RFS/SKD/baseline acc: 0.7200 for this episode
==> training...
Epoch: [1][100/288]    Time 0.121 (0.115)    Data 0.001 (0.003)    ..
Epoch: [1][200/288]    Time 0.112 (0.114)    Data 0.001 (0.002)    ...
epoch 400, total time 32.77
acc1: 0.6567, std1: 0.0076, acc2: 0.6820, std2: 0.0080,
epochs: 1, acc2: 0.6400, std2: 0.0080
...

The above is an example print-out for FC100 5-shot. acc1: 0.6567, std1: 0.0076 is the accuracy and the deviation of LinearRegression method with fixed embeddings (used in RFS and SKD). acc2: 0.6820, std2: 0.0080 is the result by our method.

Contacts

For any questions, please contact authors.

Acknowlegements

Thanks to RFS, for the preliminary implementations.

Label Hallucination for Few-Shot Classification

Related tags

Overview

Label Hallucination for Few-Shot Classification

Requirements

Download Data

Embedding Learning

Running Our Fine-tuning

Reading the outputs

Contacts

Acknowlegements

Owner

Yiren Jian

MAg: a simple learning-based patient-level aggregation method for detecting microsatellite instability from whole-slide images

Regularizing Nighttime Weirdness: Efficient Self-supervised Monocular Depth Estimation in the Dark (ICCV 2021)

Explore the Expression: Facial Expression Generation using Auxiliary Classifier Generative Adversarial Network

Codes and pretrained weights for winning submission of 2021 Brain Tumor Segmentation (BraTS) Challenge

Kaggleship: Kaggle Notebooks

Contrastive Learning of Image Representations with Cross-Video Cycle-Consistency

Train the HRNet model on ImageNet

SciPy fixes and extensions

Mixed Transformer UNet for Medical Image Segmentation

Semi-supervised Learning for Sentiment Analysis

The code for our paper Semi-Supervised Learning with Multi-Head Co-Training

Files for a tutorial to train SegNet for road scenes using the CamVid dataset

Source code for the BMVC-2021 paper "SimReg: Regression as a Simple Yet Effective Tool for Self-supervised Knowledge Distillation".

Code for our paper at ECCV 2020: Post-Training Piecewise Linear Quantization for Deep Neural Networks

📚 Papermill is a tool for parameterizing, executing, and analyzing Jupyter Notebooks.

Hierarchical probabilistic 3D U-Net, with attention mechanisms (—𝘈𝘵𝘵𝘦𝘯𝘵𝘪𝘰𝘯 𝘜-𝘕𝘦𝘵, 𝘚𝘌𝘙𝘦𝘴𝘕𝘦𝘵) and a nested decoder structure with deep supervision (—𝘜𝘕𝘦𝘵++).

Council-GAN - Implementation for our paper Breaking the Cycle - Colleagues are all you need (CVPR 2020)

Fedlearn支持前沿算法研发的Python工具库 | Fedlearn algorithm toolkit for researchers

pytorch, hand(object) detect ,yolo v5，手检测

ThunderSVM: A Fast SVM Library on GPUs and CPUs