Low-Rank Similarity Mining for Multimodal Dataset Distillation
Authors: Yue Xu, Zhilin Lin, Yusong Qiu, Cewu Lu, Yong-Lu Li
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4. Experiments, The results on Flickr30k and COCO are shown in Tab. 2 and 3. |
| Researcher Affiliation | Academia | 1Shanghai Jiao Tong University. Correspondence to: Yong-Lu Li <yonglu li@sjtu.edu.cn>. |
| Pseudocode | Yes | Algorithm 1 Low-Rank Similarity Mining (Lo RS) |
| Open Source Code | Yes | Our code is available at https: //github.com/silicx/Lo RS_Distill. |
| Open Datasets | Yes | We evaluate our method on Flickr30k (Plummer et al., 2015) and COCO (Lin et al., 2014) |
| Dataset Splits | Yes | We evaluate our method on Flickr30k (Plummer et al., 2015) and COCO (Lin et al., 2014) following the strong baseline (Wu et al., 2023) which exploits the MTT (Cazenavette et al., 2022) algorithm. The model performance is commonly measured by the recall of top K retrieval (R@K). |
| Hardware Specification | Yes | The experiments are conducted on one RTX 4090 GPU, revealing the efficiency of the method. |
| Software Dependencies | No | The paper mentions using NFNet and BERT-base models, and common deep learning frameworks are implied, but no specific software dependencies with version numbers are provided. |
| Experiment Setup | Yes | In the distillation stage, the images are resized to 224 224 and the text embeddings are 768-d. the synthetic data is learned with SGD and momentum 0.5. The rest hyperparameters including learning rate and Lo RS parameters vary among different datasets and synthetic data sizes, which are listed in Appendix Sec. F due to page limit. |