Counterfactual Reasoning for Multi-Label Image Classification via Patching-Based Training

Authors: Ming-Kun Xie, Jia-Hao Xiao, Pei Peng, Gang Niu, Masashi Sugiyama, Sheng-Jun Huang

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on multiple benchmark datasets with diverse configurations validate that the proposed method can achieve stateof-the-art performance.
Researcher Affiliation Academia 1Nanjing University of Aeronautics and Astronautics, Nanjing, China 2RIKEN Center for Advanced Intelligence Project, Tokyo, Japan 3The University of Tokyo, Tokyo, Japan.
Pseudocode Yes Algorithm 1 Pseudocode of PAT-I in a Py Torch-like style
Open Source Code Yes The implementation is available at https://github.com/xiemk/MLC-PAT.
Open Datasets Yes we conduct experiments on three benchmark datasets, including MS-COCO 2014 2 (Lin et al., 2014), Pascal VOC 2007 3 (Everingham et al., 2010), and Visual Genome 4 (Krishna et al., 2017).
Dataset Splits Yes MS-COCO contains 82,081 training images and 40,137 validation images for 80 classes, with an average of 2.9 labels per image.
Hardware Specification Yes We perform all experiments on Ge Force RTX 3090 GPUs.
Software Dependencies No The paper mentions using the Adam optimizer and implementing in a 'Py Torch-like style,' but it does not specify exact version numbers for any software libraries or frameworks, such as PyTorch version.
Experiment Setup Yes We employ Adam (Kingma & Ba, 2014) optimizer and one-cycle policy scheduler to train the model with maximal learning rate of 0.0001. Furthermore, we perform exponential moving average (EMA) (Tarvainen & Valpola, 2017) for the model parameters with a decay of 0.9997. The random seed is set to 1 for all experiments.