Differentiable Dynamic Normalization for Learning Deep Representation

Authors: Ping Luo, Peng Zhanglin, Shao Wenqi, Zhang Ruimao, Ren Jiamin, Wu Lingyun

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive studies show that DN outperforms its counterparts in CIFAR10 and Image Net. and Sec.4.1 evaluates DN in CIFAR10 (Krizhevsky, 2009) and Image Net (Russakovsky et al., 2015), where it is demonstrated by comparing with previous normalization techniques. Ablation studies are presented in Sec.4.3.
Researcher Affiliation Collaboration 1Department of Computer Science, The University of Hong Kong 2Department of Electronic Engineering, The Chinese University of Hong Kong 3Sense Time Group Ltd.
Pseudocode Yes Algorithm 1 Computations of DN
Open Source Code No The paper does not provide any specific repository link or explicit statement about the availability of the source code for the methodology described.
Open Datasets Yes Extensive studies in CIFAR10 (Krizhevsky, 2009) and Image Net (Russakovsky et al., 2015) demonstrate that DN is able to outperform its counterparts.
Dataset Splits No The paper mentions evaluating on CIFAR10 and ImageNet, and reports results on the 'validation set' for ImageNet, but does not explicitly provide specific percentages, sample counts, or clear statements about the training, validation, and test splits used for its experiments needed for reproduction.
Hardware Specification No The paper mentions that gradients are 'aggregated across GPUs' but does not specify any particular GPU model, CPU, or other hardware used for running the experiments.
Software Dependencies No The paper mentions 'Tensor Flow and Py Torch' as platforms for implementation but does not specify any software dependencies with version numbers.
Experiment Setup Yes All models are trained on CIFAR10 with different batch sizes, where the gradients are aggregated across GPUs, while the statistics are estimated within each GPU. We repeat to train all models five times... and The batch size is (8, 32).