Research map / Detection and segmentation / CNN features
Fast Neural Network Adaptation via Parameter Remapping and Architecture Search
Jiemin Fang*, Yuzhu Sun*, Kangjian Peng*, Qian Zhang, Yuan Li, Wenyu Liu, Xinggang Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Adapt MobileNetV2 parameters and architecture with FNA | MobileNetV2 in DeepLabv3 on Cityscapes · Cityscapes | mIOU (%): 75.5 -> 76.6 | Improved |
| Adapt MobileNetV2 with FNA | RetinaNet on MS-COCO · MS-COCO | mAP (%): 33.3 -> 33.9 | Improved |
| Adapt MobileNetV2 with FNA | SSDLite on MS-COCO · MS-COCO | mAP (%): 22.1 -> 23.3 | Improved |
| Add parameter remapping before architecture adaptation (ablation) | FNA on Cityscapes with DeepLabv3 · Cityscapes | mIOU (%): 76.0 -> 76.6 | Improved |
| Add parameter remapping before parameter adaptation (ablation) | FNA on Cityscapes with DeepLabv3 · Cityscapes | mIOU (%): 73.0 -> 76.6 | Improved |
| Add remapping before both architecture and parameter adaptation (ablation) | FNA on Cityscapes with DeepLabv3 · Cityscapes | mIOU (%): 72.4 -> 76.6 | Improved |
| Use remapping rather than ImageNet pretraining before parameter adaptation (ablation) | FNA on Cityscapes with DeepLabv3 · Cityscapes | mIOU (%): 76.5 -> 76.6 | Improved |
| Replace differentiable search with random search (ablation) | FNA on RetinaNet · MS-COCO | mAP (%): 33.9 -> 33.5 | Hurt |
| Add parameter remapping before architecture search (ablation) | Random-search FNA on RetinaNet · MS-COCO | mAP (%): 31.5 -> 33.5 | Improved |
| Add parameter remapping before parameter adaptation (ablation) | Random-search FNA on RetinaNet · MS-COCO | mAP (%): 25.3 -> 33.5 | Improved |
| Select width channels by batch-normalization statistics rather than FNA's basic remapping (ablation) | Parameter remapping on Cityscapes with DeepLabv3 · Cityscapes | mIOU (%): 76.6 -> 75.8 | Hurt |
| Select width channels by weight standard deviation rather than FNA's basic remapping (ablation) | Parameter remapping on Cityscapes with DeepLabv3 · Cityscapes | mIOU (%): 76.6 -> 75.8 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MobileNetV2 — MobileNetV2: Inverted Residuals and Linear Bottlenecks (CVPR 2018) · better on 3 of 3 shared results
- DetNAS — DetNAS: Backbone Search for Object Detection (NeurIPS 2019) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FNA++: Fast Network Adaptation via Parameter Remapping and Architecture Search (TPAMI 2020 · similarity 0.97)
- Auto-FPN: Automatic Network Architecture Adaptation for Object Detection Beyond Classification (ICCV 2019 · similarity 0.91)
- Representation Sharing for Fast Object Detector Search and Beyond (ECCV 2020 · similarity 0.89)
- DIY Your EasyNAS for Vision: Convolution Operation Merging, Map Channel Reducing, and Search Space to Supernet Conversion Tooling (TPAMI 2023 · similarity 0.89)
- NAS-FCOS: Fast Neural Architecture Search for Object Detection (CVPR 2020 · similarity 0.88)
- Neural Architecture Search for Dense Prediction Tasks in Computer Vision (IJCV 2023 · similarity 0.88)
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.
