A Lower Bound of Hash Codes' Performance
Authors: Xiaosu Zhu, Jingkuan Song, Yu Lei, Lianli Gao, Hengtao Shen
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments reveal the effectiveness of the proposed method. By testing on a series of hash-models, we obtain performance improvements among all of them, with an up to 26.5% increase in mean Average Precision and an up to 20.5% increase in accuracy. |
| Researcher Affiliation | Academia | 1Center for Future Media, University of Electronic Science and Technology of China 2Peng Cheng Laboratory |
| Pseudocode | Yes | Algorithm 1 One of implementations under supervised circumstance. |
| Open Source Code | Yes | Our code is publicly available at https://github.com/VL-Group/LBHash. |
| Open Datasets | Yes | CIFAR-10 [20] is a single-label 10-class dataset. The whole dataset contains 6,000 images for each class. Following [5], we split the dataset into 500 | 5,400 | 100 for each class randomly as train, base and query splits, respectively. NUS-WIDE [7] consists of 81 labels and images may have one or more labels. We follow previous works [5] to pick the most frequent 21 labels and their associated images (195,834) for experiments. Specifically, 193,734 images are randomly picked to form the base split while remaining 2,100 images are adopted for queries. 10,500 images are randomly sampled from the base split for training models. Image Net [11] is a large-scale dataset consists of 1,000 classes. To conduct experiments, we follow [5] to pick a subset of 100 classes where all images of these classes in the training set / validation set are as base split / query split respectively (128,503 | 4,983 images). We then randomly sample 100 images per class in the base split for training. |
| Dataset Splits | Yes | Following [5], we split the dataset into 500 | 5,400 | 100 for each class randomly as train, base and query splits, respectively. Specifically, 193,734 images are randomly picked to form the base split while remaining 2,100 images are adopted for queries. 10,500 images are randomly sampled from the base split for training models. To conduct experiments, we follow [5] to pick a subset of 100 classes where all images of these classes in the training set / validation set are as base split / query split respectively (128,503 | 4,983 images). We then randomly sample 100 images per class in the base split for training. |
| Hardware Specification | Yes | We also measure averaged training time per epoch of all variants on a single NVIDIA RTX 3090, which is placed in Tab. 1. |
| Software Dependencies | No | The paper mentions 'PyTorch [34]' but does not specify a version number or other software dependencies with versions. |
| Experiment Setup | Yes | We adopt Adam [35] optimizer with default configuration and learning rate of our method η1 = η2 = 1e-3 for training. For multi-label datasets, we simply modify Alg. 1 Line 8 with the sum of multiple losses. Block number u of P is set to bits/8. For example, if the length of hash code is 64, there will be 8 sub-models Pπ1 Pπ8 trained in parallel. |