Research map / Detection and segmentation / CNN features
Bag of Tricks for Image Classification with Convolutional Neural Networks
Check 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 |
|---|---|---|---|
| Increase batch size from 256 to 1024 using linear learning-rate scaling alone (ablation) | ResNet-50 · ImageNet validation | top-1 accuracy: 75.87 -> 75.17 | Hurt |
| Add learning-rate warmup to linear scaling (ablation) | ResNet-50 · ImageNet validation | top-1 accuracy: 75.17 -> 75.93 | Helped |
| Initialize the final batch-normalization gamma in residual blocks to zero (ablation) | ResNet-50 · ImageNet validation | top-1 accuracy: 75.93 -> 76.37 | Helped |
| Exclude biases and batch-normalization parameters from weight decay (ablation) | ResNet-50 · ImageNet validation | top-1 accuracy: 76.37 -> 76.03 | Hurt |
| Switch to FP16 at the end of training (ablation) | ResNet-50 · ImageNet validation | top-1 accuracy: 76.03 -> 76.21 | No clear effect |
| Use ResNet-B stride placement in the downsampling block (ablation) | ResNet-50 · ImageNet validation | top-1 accuracy: 76.21 -> 76.66 | Helped |
| Replace the stem 7x7 convolution with three 3x3 convolutions (ablation) | ResNet-50-B · ImageNet validation | top-1 accuracy: 76.66 -> 76.87 | Helped |
| Add average pooling before the downsampling path-B convolution (ablation) | ResNet-50-C · ImageNet validation | top-1 accuracy: 76.87 -> 77.16 | Helped |
| Replace step decay with cosine learning-rate decay (ablation) | ResNet-50-D · ImageNet validation | top-1 accuracy: 77.16 -> 77.91 | Helped |
| Add label smoothing after cosine decay (ablation) | ResNet-50-D · ImageNet validation | top-1 accuracy: 77.91 -> 78.31 | Helped |
| Add distillation without mixup (ablation) | ResNet-50-D · ImageNet validation | top-1 accuracy: 78.31 -> 78.67 | Helped |
| Add mixup without distillation (ablation) | ResNet-50-D · ImageNet validation | top-1 accuracy: 78.31 -> 79.15 | Helped |
Closest papers by meaning
- Do Better ImageNet Models Transfer Better? (CVPR 2019 · similarity 0.89)
- When does dough become a bagel? Analyzing the remaining mistakes on ImageNet (NeurIPS 2022 · similarity 0.87)
- Does progress on ImageNet transfer to real-world datasets? (NeurIPS 2023 · similarity 0.87)
- An Empirical Evaluation of Current Convolutional Architectures' Ability to Manage Nuisance Location and Scale Variability (CVPR 2016 · similarity 0.86)
- ImageNet-Hard: The Hardest Images Remaining from a Study of the Power of Zoom and Spatial Biases in Image Classification (NeurIPS 2023 · similarity 0.86)
- Doomed to Re-Annotate, Forever: The ImageNet Story (NeurIPS 2026 · similarity 0.86)
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.