Research map / Architectures and efficiency / RNNs and normalization
In-Place Activated BatchNorm for Memory-Optimized Training of DNNs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace standard BN+activation with InPlace-ABN and use the freed memory for a larger training batch | ResNeXt-101 · ImageNet-1k val, 224² center | top-1 accuracy: 77.04 -> 78.08 | Helped |
| Use the freed memory to train a deeper or wider model | ImageNet-1k classification · ImageNet-1k val, 224² center | top-1 accuracy: 77.04 -> 78.28; 77.04 -> 79.72 | Helped |
| Replace ReLU with Leaky ReLU at validation for a ReLU-trained ResNeXt-101 (ablation) | ResNeXt-101 · ImageNet-1k val, 224² center | top-1 accuracy: 77.74 -> 76.88 | Hurt |
| Replace Leaky ReLU with ReLU at validation for a Leaky-ReLU-trained ResNeXt-101 (ablation) | ResNeXt-101 · ImageNet-1k val, 224² center | top-1 accuracy: 77.04 -> 76.81 | No clear effect |
| Use InPlace-ABN to fit more training crops per minibatch | ResNeXt-101 semantic segmentation · Cityscapes val | score (%): 74.42 -> 75.80 | Helped |
| Use InPlace-ABN to increase training crop resolution at fixed batch size | WideResNet-38 semantic segmentation · Cityscapes val | score (%): 75.82 -> 78.31 | Helped |
| Synchronize BatchNorm statistics across GPUs (ablation) | ResNeXt-101 classification · ImageNet-1k val, 224² center | top-1 accuracy: 77.04 -> 77.70 | Helped |
| Add Class-Uniform sampling (ablation) | WideResNet-38 with synchronized InPlace-ABN · Cityscapes val | score (%): 79.16 -> 79.40 | Helped |
| Increase crop size from 712² to 872² with synchronized InPlace-ABN (ablation) | WideResNet-38 · Cityscapes val | score (%): 79.02 -> 79.16 | Helped |
| Replace standard BN+activation with InPlace-ABN | Training memory for modern DNNs · Training | memory reduction: up to 50% | Helped |
| Use InPlace-ABN instead of Standard execution | ResNeXt-101 BN+Act+Conv blocks · Timing analysis | computation time increase: 0.8-2% | Hurt |
Closest papers by meaning
- Deconstructing the Regularization of BatchNorm (ICLR 2021 · similarity 0.88)
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.88)
- Beyond BatchNorm: Towards a Unified Understanding of Normalization in Deep Learning (NeurIPS 2021 · similarity 0.88)
- Four Things Everyone Should Know to Improve Batch Normalization (ICLR 2020 · similarity 0.87)
- Filter Response Normalization Layer: Eliminating Batch Dependence in the Training of Deep Neural Networks (CVPR 2020 · similarity 0.87)
- Reducing Divergence in Batch Normalization for Domain Adaptation (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.