Explore Visual Concept Formation for Image Classification
Authors: Shengzhou Xiong, Yihua Tan, Guoyou Wang
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that LSOVCF improves the ability of identifying unseen samples on cifar10, STL10, flower17 and Image Net based on several backbones, from the classic VGG to the SOTA Ghostnet. |
| Researcher Affiliation | Academia | 1National Key Laboratory of Science & Technology on Multispectral Information Processing, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074 , China. |
| Pseudocode | Yes | Algorithm 1 Learning Strategy of Concept Formation |
| Open Source Code | Yes | The code is available at https: //github.com/elvintanhust/LSOVCF. |
| Open Datasets | Yes | Experiments are conducted on four datasets, including cifar10 (Krizhevsky, 2012), STL10 (Coates et al., 2011), flower17 (Nilsback & Zisserman, 2006) and a subset of Image Net (Russakovsky et al., 2015). |
| Dataset Splits | Yes | Cifar10 dataset consists of 60000 images in 10 classes, with 6000 images per class and 5000 of them are for training. And in our experiments, we randomly select 500 images from each class s training set for validation. STL10 dataset have 1300 labeled images per class, for each class, the training set, test set and validation set are divided with a ratio of 750:500:50. Flower17 dataset consists of 1360 labeled images that belong to 17 kinds of flowers, we take 70 images of each class as training set and 10 image as test set, there is no validation set in flower17 dataset. As for Imagenet, 100 categories are randomly selected for experiments because the time consumption on the complete dataset is too high. And 100 samples of each category are randomly selected as test set. |
| Hardware Specification | Yes | Finally, all experiments are based on the Py Torch and performed on RTX 2080 Ti GPU. |
| Software Dependencies | No | Finally, all experiments are based on the Py Torch and performed on RTX 2080 Ti GPU. |
| Experiment Setup | Yes | The capacity and sampling batch size of ECP are set to (1500, 500) for cifar10, (600, 200) for STL10, (300, 100) for flower17 and (1300, 500) for Image Net. And the training batch size of cifar10, STL10, flower17 and Image Net are set to 100, 100, 170, 500. ... All models will be optimized by stochastic gradient descent and the initial learning rate is set to 0.1, except that learning rate of VGG is set to 0.01 for convergence. In addition, all models will be trained 200 epochs with the same data preprocessing, and the learning rate will decay by a factor of 10 in 100th and 150th epoch. |