AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Deeply Tensor Compressed Transformers for End-to-End Object Detection

Peining Zhen, Ziyang Gao, Tianshu Hou, Yuan Cheng, Hai-Bao Chen

AAAI 2022 · Architectures and efficiency › Attention and transformers · approach: Efficient image transformers · 12 citations (OpenAlex)

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 changedWhereMeasuredEffect
Tensor-train compress FFN weights and add GMHADETR with ResNet-50 · COCO valAP: 40.1 -> 39.5Hurt
Add GMHA to vanilla DETR (ablation)DETR · COCO valAP: 40.1 -> 41.7Improved
Add GMHA to tensorized DETR (ablation)T-DETR · COCO valAP: 37.6 -> 39.5Improved
Apply GMHA only to encoder self-attention rather than all attention modules (ablation)T-DETR · COCO valAP: 39.2 -> 39.5Improved
Increase tensor-train rank from 3 to 4 (ablation)Tensorized DETR FFN · COCO valAP: 36.9 -> 37.6Improved
Increase tensor-train rank from 4 to 5 (ablation)Tensorized DETR FFN · COCO valAP: 37.6 -> 37.7Improved
Replace tensor-train decomposition with Tucker decomposition (ablation)DETR FFN · COCO valAP: Tucker rank 2: 37.4; rank 3: 37.7; Tucker-2 rank 4: 37.6Mixed
Quantize the ResNet-50 backbone to 8-bitT-DETR · COCO valAP: 39.5 -> 39.5No clear effect
Quantize the ResNet-50 backbone to 4-bitT-DETR · COCO valAP: 39.5 -> 37.9Hurt
Quantize ResNet-50 weights to 8-bitResNet-50 · ImageNet-1k valtop-1 accuracy: 76.15 -> 76.18No clear effect
Quantize ResNet-50 weights to 4-bitResNet-50 · ImageNet-1k valtop-1 accuracy: 76.15 -> 75.51Hurt
Use tensor-train rank 4 rather than an uncompressed FFN (ablation)DETR FFN · COCO valFFN parameter count: 1,048,576 -> 2,176Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy