Label-Aware Global Consistency for Multi-Label Learning with Single Positive Labels
Authors: Ming-Kun Xie, Jiahao Xiao, Sheng-Jun Huang
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on multiple benchmark datasets validate that the proposed method can achieve state-of-the-art performance for solving SPML tasks. In this section, we first perform experiments to compare our method with state-of-the-art methods; then, we conduct ablation studies to examine the effectiveness of each component for the proposed method. |
| Researcher Affiliation | Academia | Ming-Kun Xie , Jia-Hao Xiao , Sheng-Jun Huang College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics MIIT Key Laboratory of Pattern Analysis and Machine Intelligence, Nanjing, 211106 {mkxie,jiahaoxiao,huangsj}@nuaa.edu.cn |
| Pseudocode | No | The paper describes the proposed method using equations and textual descriptions, but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement or link for the open-sourcing of the code for their proposed methodology. It only mentions using code from a prior work [4] for dataset generation. |
| Open Datasets | Yes | We perform experiments to evaluate our proposed method on four benchmark datasets: Pascal VOC-2012 (VOC for short) 3 [10], MS-COCO-2014 (COCO for short) 4 [17], NUS-WIDE (NUS for short) 5 [3], and CUB-200-2011 (CUB for short) 6 [27]. |
| Dataset Splits | Yes | Specifically, VOC contains 5,717 training images and 5,823 validation images for 20 classes. We divide the training set into 4,574 training examples and 1,143 validation examples, and then use the original validation set for testing. COCO contains 82,081 training images and 40,137 validation images for 80 classes. We divide the training set into 65,665 training examples and 16,416 validation examples, and then use the original validation set for testing. NUS...we merge all images and randomly select 150,000 training samples and 60,260 testing examples. Finally, we withhold 30,000 images from training examples for validation. CUB...We divide the original training set into 4,795 training examples and 1,199 validation examples. |
| Hardware Specification | Yes | We perform all experiments on Ge Force RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions software components like AdamW optimizer, One Cycle LR scheduler, ResNet-50, Cutout, and Rand Augment, but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | We resize the resolution of input images to 448 448 as in [4]. For training images, we use both a weak augmentation (only containing random horizontal flipping) and a strong augmentation (containing Cutout [8] and Rand Augment [5]). For training the model, we use the Adam W [19] optimizer with the weight decay of 0.01. The One Cycle LR scheduler is used to change the learning rate with the max learning rate of 0.0001. We train the model for 40 epochs with the early stopping. We consider the batch size in the range of {8, 16, 32, 64}. At the first stage, we set λPCL = 1 and determine the threshold δ from the range of {0.5, 0.6, 0.7, 0.8, 0.9}. At the second stage, there are two extra parameters, including the balancing parameter λLAC, the size of memory queue K. We set λLAC = 1, and determine K from the range of {512, 1024, 2048, 4096, 8192}, respectively. All hyperparameters are determined according to their m AP obtained on the validation set. Furthermore, we apply exponential moving average (EMA) to model parameters θ with a decay of 0.9997. The random seed is set to 1 for all experiments. |