Inconsistency-Based Data-Centric Active Open-Set Annotation

An implementation of the NEAT batch active learning algorithm for Open-Set Annotation. Details are provided in our paper: https://ojs.aaai.org/index.php/AAAI/article/view/28213

1. Requirements

Environments

Currently, requires following packages. (We are using CUDA == 11.6, python == 3.9, pytorch == 1.13.0, torchvision == 0.14.0, scikit-learn == 1.2.2, matplotlib == 3.7.1, numpy == 1.23.5)

• CUDA 10.1+
• python 3.7.9+
• pytorch 1.7.1+
• torchvision 0.8.2+
• scikit-learn 0.24.0+
• matplotlib 3.3.3+
• numpy 1.19.2+

Datasets

For CIFAR10 and CIFAR100, we provide a function to automatically download and preprocess the data, you can also download the datasets from the link, and please download it to ~/data folder.

• CIFAR10
• CIFAR100
• TinyImagenet

2. Get started

$ cd Active-OpenSet-NEAT

Although We have provided scripts to automatically download CIFAR10 and CIFAR100 dataset but for Tiny-Imagenet or Imagenet you are supposed to download yourself utilizing the following command or using the link to download manually.

$ mkdir data
$ cd data
$ wget http://cs231n.stanford.edu/tiny-imagenet-200.zip
$ unzip tiny-imagenet-200.zip

After you have all the dataset available you also need to run the extract_features.py to extract features using CLIP for all dataset if you want to specify certain dataset please change the script.

$ mkdir features
$ python extract_features.py

3. Training all the active learning strategies mentioned in our paper

• run the following command in the terminal(example).
• You have the freedom to adjust the arguments according to your interest in modifying the command, depending on the "Option" provided below.

$ python NEAT_main.py --gpu 1 --k 10 --save-dir log_AL/ --query-strategy NEAT --init-percent 1 --known-class 2 --query-batch 400 --seed 2 --model resnet18 --dataset cifar10

• Option
• --datatset: cifar10, cifar100, and Tiny-Imagenet.
• --known-class: 2, 20, and 40 for cifar10, cifar100, Tiny-Imagenet respectively in our experiments.
• --init-percent: 1, 8, 8 for cifar10, cifar100, Tiny-Imagenet respectively in our experiments.
• --query-batch 400 in our experiments
• --model: 'resnet18', 'resnet34', 'resnet50', and 'vgg16'
• --query-strategy: 'random', 'uncertainty', 'AV_temperature', 'NEAT_passive', 'NEAT', "BGADL", "OpenMax", "Core_set", 'BADGE_sampling', "certainty", "hybrid-BGADL", "hybrid-OpenMax", "hybrid-Core_set", "hybrid-BADGE_sampling", "hybrid-uncertainty"
• --workers: default 4 in our setup if you only have one gpu please set to 0.
• --max-epoch: 100 in our experiments.
• --max-query: 10 in our experiments.
• --k: 10 number of neighbors you can change this number based on your research requirements.
• --pre-type: default is clip, and you can introduce your pre-trained model base on your interest.

4. Evaluation

To evaluate the performance of NEAT, we provide a set of plotting python scripts.

• Option
• the following jupyter notebook will plot Accuracy, Precision, and Recall for all the active learning methods mentioned in our paper.

$ plot_baseline_batch_size_400.ipynb

• the following python file can plot Accuracy, Precision, and Recall for different number of neighbors.

$ plot_different_k.py

• the following python file can plot Accuracy, Precision, and Recall for different pre-trained models.

$ plot_pretrained_model.py

• the following python file can plot the Accuracy, Precision, and Recall for all active learning methods with batch size of 600 and 800.

$ plot_batch_600_and_800.py

• the following python file can plot the T-sne representation of comparison of NEAT and LFOSA.

$ T-sne_plot_NEAT_and_LFOSA.py