Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Quantization Robust Federated Learning for Efficient Inference on Heterogeneous Devices
Authors: Kartik Gupta, Marios Fournarakis, Matthias Reisser, Christos Louizos, Markus Nagel
TMLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments on standard FL benchmarks to evaluate our proposed Fed Avg variants for quantization robustness and provide a convergence analysis for our Quantization-Aware variants in FL. and We perform extensive experimental evaluations of baselines and our Fed Avg variants on CIFAR-10, CIFAR-100, FEMNIST and Tiny Image Net with different network architectures. |
| Researcher Affiliation | Collaboration | Australian National University, Qualcomm AI Research EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Fed Avg , Fed Avg-KURE and Algorithm 2 Fed Avg-APQN , Fed Avg-QAT , and Fed Avg-MQAT |
| Open Source Code | No | The paper does not provide an explicit statement about the release of source code for the methodology described, nor does it include a direct link to a code repository. |
| Open Datasets | Yes | For the experimental comparisons, we use federated versions of the CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009), Tiny Image Net1 and FEMNIST (Caldas et al., 2018a) datasets. and 1https://tiny-imagenet.herokuapp.com/ |
| Dataset Splits | Yes | We split the data into 100 (CIFAR-10), 500 (CIFAR-100), 500 (Tiny Image Net), 3500 (FEMNIST) clients in a non-i.i.d way following Hsu et al. (2019), where in each round only 10 clients participate for all datasets except Tiny Image Net dataset, where 100 clients participate. and In order to simulate a non-i.i.d. data split that we would expect in the federated scenario, we artificially split CIFAR-10, CIFAR-100 and Tiny Image Net by their label. For CIFAR-10 and Tiny Image Net, the label proportions on each client are computed by sampling from a Dirichlet distribution with α 1.0 (Hsu et al. (2019)). For CIFAR-100 we use the coarse labels provided with the dataset and follow Reddi et al. (2020). For our FEMNIST experiments, the federated split is naturally determined by the writer-id for each client. |
| Hardware Specification | Yes | Our experiments are performed using NVIDIA Tesla V100 GPUs and code is in Py Torch. |
| Software Dependencies | No | The paper mentions 'Py Torch' as the framework used, but does not specify any version numbers for PyTorch or other critical libraries. |
| Experiment Setup | Yes | For all our experiments, we use ADAM optimizer for server training phase and SGD optimizer for client training phase. We use single epoch of local client training for each client participating in a round for all our experiments. and Table 2: Hyperparameters used for the experimental evaluations in the paper. Here ηs, ηc denote the server and client learning rate and ϵs refers to the correction term in value in ADAM optimizer of server. |