Batch-Instance Normalization for Adaptively Style-Invariant Neural Networks

Authors: Hyeonseob Nam, Hyo-Eun Kim

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments show that the proposed approach surprisingly outperforms BN in general object classification and multi-domain problems, and it also successfully substitutes IN in style transfer networks.
Researcher Affiliation Industry Hyeonseob Nam Lunit Inc. Seoul, South Korea hsnam@lunit.ioHyo-Eun Kim Lunit Inc. Seoul, South Korea hekim@lunit.io
Pseudocode No The paper provides mathematical equations for Batch-Instance Normalization but does not include formal pseudocode blocks or algorithms.
Open Source Code Yes BIN can be implemented with only a few lines of code using popular deep learning frameworks.1; 1https://github.com/hyeonseob-nam/Batch-Instance-Normalization
Open Datasets Yes We first evaluate BIN for general object classification using CIFAR-10/100 [15] and Image Net [18] datasets. ... We employ the Office-Home dataset [23] ... using content images from the MS-COCO dataset [17]
Dataset Splits Yes Table 1: Top-1 accuracy (%) on CIFAR-10/100 (test set) and Image Net (validation set) evaluated with Res Net-110 and Res Net-18, respectively. ... On CIFAR datasets, we train the networks for 64K iterations, where the learning rate is divided by 10 at 32K and 48K iterations; on Image Net, training runs for 90 epochs, where the learning rate is divided by 10 at 30 and 60 epochs.
Hardware Specification No The paper describes the network architectures and training procedures but does not specify the hardware (e.g., GPU models, CPU types) used for experiments.
Software Dependencies No The paper mentions 'popular deep learning frameworks' and 'publicly available implementation2 of classification networks' (footnote 2 links to 'pytorch-classification') but does not specify software dependencies with version numbers.
Experiment Setup Yes The networks are trained by SGD with a batch size of 128, an initial learning rate of 0.1, a momentum of 0.9, and a weight decay of 0.0001. On CIFAR datasets, we train the networks for 64K iterations, where the learning rate is divided by 10 at 32K and 48K iterations; on Image Net, training runs for 90 epochs, where the learning rate is divided by 10 at 30 and 60 epochs.