Language-Guided Transformer for Federated Multi-Label Classification

Authors: I-Jieh Liu, Ci-Siang Lin, Fu-En Yang, Yu-Chiang Frank Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments on various multi-label datasets (e.g., FLAIR, MSCOCO, etc.), we show that our Fed LGT is able to achieve satisfactory performance and outperforms standard FL techniques under multi-label FL scenarios. Code is available at https://github.com/Jack24658735/Fed LGT.
Researcher Affiliation Collaboration I-Jieh Liu1, Ci-Siang Lin1,2, Fu-En Yang1,2, Yu-Chiang Frank Wang1,2 1Graduate Institute of Communication Engineering, National Taiwan University 2NVIDIA
Pseudocode Yes The pseudo-code of our proposed framework is described in Algorithm 1.
Open Source Code Yes Code is available at https://github.com/Jack24658735/Fed LGT.
Open Datasets Yes we conduct extensive evaluations on various benchmark datasets, including FLAIR (Song, Granqvist, and Talwar 2022), MS-COCO (Lin et al. 2014), and PASCAL VOC (Everingham et al. 2015).
Dataset Splits No The paper does not provide specific percentages or counts for training, validation, and test splits for the datasets. It mentions 'FLAIR provides real-user data partitions' but no details on the splits.
Hardware Specification Yes For all experiments, we implement our model by Py Torch and conduct training on a single NVIDIA RTX 3090Ti GPU with 24GB memory.
Software Dependencies No The paper mentions 'Py Torch' as the implementation framework and 'Adam' as the optimizer, but does not specify their version numbers or other software dependencies with versions.
Experiment Setup Yes set the threshold τ to 0.5 and the uncertainty margin ε is 0.02. For each round of local training, we train 5 epochs using the Adam (Kingma and Ba 2014) optimizer with a learning rate of 0.0001, and the batch size is set to 16. For the detail settings about FL, the communication round T is set to 50, and the fraction of active clients in each round is designed to achieve a level of participation equivalent to 50 clients, thus ensuring the data distribution is representative of the overall population.