├── LICENSE
├── README.md
├── cga.py
├── configs
├── deit_default_imagent.attn_q.yml
├── ours_imagenet_recipe.attn_q.yml
└── swin_t_imagenet.attn_q.yml
├── eval.py
├── eval_scripts
├── deit_s
│ ├── w2a2.sh
│ ├── w3a3.sh
│ └── w4a4.sh
├── deit_t
│ ├── w2a2.sh
│ ├── w3a3.sh
│ └── w4a4.sh
└── swin_t
│ ├── w2a2.sh
│ ├── w3a3.sh
│ └── w4a4.sh
├── imgs
├── ofq_vit.png
└── w2a2_deit_s_cga_vis_.png
├── src
├── __init__.py
├── deit.py
├── deit_vision_transformer.py
├── quantization
│ ├── __init__.py
│ ├── modules
│ │ ├── __init__.py
│ │ ├── attention.py
│ │ ├── qbias.py
│ │ ├── qlinear.py
│ │ ├── swin_attention_and_mlp.py
│ │ └── utils.py
│ ├── quantizer
│ │ ├── __init__.py
│ │ ├── lsq.py
│ │ └── statsq.py
│ └── utils.py
├── registry.py
├── swin.py
└── utils
│ ├── __init__.py
│ ├── embedder.py
│ ├── helpers.py
│ ├── stochastic_depth.py
│ ├── tokenizer.py
│ ├── transformers.py
│ └── utils.py
├── timm_fix_imagenet_loading_bugs
└── dataset_factory.py
├── train.py
└── train_scripts
├── deit_s
├── w2a2_deit_s.sh
├── w3a3_deit_s.sh
└── w4a4_deit_s.sh
├── deit_t
├── w2a2_deit_t.sh
├── w3a3_deit_t.sh
└── w4a4_deit_t.sh
└── swin_t
├── w2a2_swin_t.sh
├── w3a3_swin_t.sh
└── w4a4_swin_t.sh
/LICENSE:
--------------------------------------------------------------------------------
1 | MIT License
2 |
3 | Copyright (c) 2023 LIU, Shih-yang
4 |
5 | Permission is hereby granted, free of charge, to any person obtaining a copy
6 | of this software and associated documentation files (the "Software"), to deal
7 | in the Software without restriction, including without limitation the rights
8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9 | copies of the Software, and to permit persons to whom the Software is
10 | furnished to do so, subject to the following conditions:
11 |
12 | The above copyright notice and this permission notice shall be included in all
13 | copies or substantial portions of the Software.
14 |
15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 | SOFTWARE.
22 |
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # OFQ: Oscillation-free Quantization for Low-bit Vision Transformers
2 |
3 | This repository contains the training code of ViT introduced in our work: "[Oscillation-free Quantization for Low-bit Vision Transformers](https://arxiv.org/abs/2302.02210)" which has been accepted for ICML 2023. Please consider starring the repo if you find our work useful, thanks!
4 |
5 | In this work, we discusses the issue of weight oscillation in quantization-aware training and how it negatively affects model performance. The learnable scaling factor, commonly used in quantization, was found to worsen weight oscillation. The study proposes three techniques to address this issue: statistical weight quantization (StatsQ), confidence-guided annealing (CGA), and query-key reparameterization (QKR). These techniques were tested on the ViT model and were found to improve quantization robustness and accuracy. The proposed 2-bit DeiT-T/DeiT-S algorithms outperform the previous state-of-the-art by 9.8% and 7.7%, respectively.
6 |
7 |
8 |
9 |
10 |
Fig.1 - Trajectory of statistical scaling factors (StatsQ) from the 10th transformer blocks in a 2-bit DeiT-S throughout CGA with 4 different boundary ranges [BR_0.003, BR_0.005, BR_0.007, BR_0.01]. The y-axis represents the value of scaling factorsß
11 |