Welcome to On-Device AI: ON THE AIr repository! We aim to research On-Device AI with an emphasis on common model compression techniques, conducting paper reviews, and benchmarking real-world performance using NVIDIA Jetson devices. Join us in advancing On-Device AI through open collaboration and innovation! ๐
"Propose the optimal model compression techniques for NVIDIA Jetson devices by leveraging the knowledge gained from research paper reviews on model compression methods."
- Learn various pruning techniques during this season (10th).
- Apply the learned model compression methods to existing models.
- Test the actual performance on the NVIDIA Jetson platform.
- Share the results for collaborative insights and community contribution.
- Foster synergy between individual growth and collective intelligence.
- Promote a knowledge-sharing culture based on the open-source spirit.
์ญํ | ์ด๋ฆ | ๊ธฐ์ ์คํ ๋ฐฐ์ง | ์ฃผ์ ๊ด์ฌ ๋ถ์ผ |
---|---|---|---|
Project Manager | ์ ํ์ฐ | On-Device AI, CV, Robotics | |
Member | ๊น๋ฏผ์ฑ | - | |
Member | ๊ตฌ์น์ฐ | - | |
Member | ๋ฌธ๊ท์ | - | |
Member | ๋ฐ์ ์ | - | |
Member | ๋ฐ์๋ฆฌ | - | |
Member | ์๋ฌธ๊ธฐ | - | |
Member | ์ ์ง์ฐ | - | |
Member | ์ต์์ | - | |
Member | ์ต์ ์ง | - | |
Member | ์ตํด์ธ | - |
gantt
title 2025 On-Device AI ํ๋ก์ ํธ ์ฌ์
section ์ ์ฒด ์ปค๋ฆฌํ๋ผ
Pruning :a1, 2025-03-03, 119d
Quantization :a2, after a1, 120d
section Pruning ์ธ๋ถ ํ๋
SPECIFIC OR UNIVERSAL SPEEDUP :b1, 2025-03-03, 35d
WHEN TO PRUNE :b2, after b1, 84d
section ์ค์ต ์ธ๋ถ ํ๋ with Jetson
Object Detection with Pruning :c1, 2025-04-01, 63d
LLM with Pruning :c2, after c1, 30d
CV with Pruning :c3, after c1, 30d
๋ ์ง | ๋ด์ฉ | ๋ฐํ์ | ์งํ๋ฐฉ์ | ์ฐธ๊ณ ์๋ฃ | ๋น๊ณ |
---|---|---|---|---|---|
2025/03/05 | OT | ์ ํ์ฐ | ์จ๋ผ์ธ | - | |
2025/03/12 | Unstructured Pruning | ๊ตฌ์น์ฐ | ์จ๋ผ์ธ | J. Frankle and M. Carbin, โThe lottery ticket hypothesis: finding sparse, trainable neural networks,โ in ICLR, 2019. | |
2025/03/19 | Structured Pruning | ๊น๋ฏผ์ฑ | ์คํ๋ผ์ธ | X. Ma, G. Fang, and X. Wang, โLLM-Pruner: On the structural pruning of large language models,โ in NeurIPS, vol. 36, 2023, pp.21 702โ21 720. | |
2025/03/26 | Magical Week ํด์ผ | ๋ฏธ์ | - | - | |
2025/04/03 | Semi-structured Pruning | ์ต์ ์ง | ์จ๋ผ์ธ | F. Meng, H. Cheng, K. Li, H. Luo, X. Guo, G. Lu, and X. Sun, โPruning filter in filter,โ in NeurIPSW, 2020. | |
2025/04/09 | Pruning Before Training | ๋ฌธ๊ท์ | ์จ๋ผ์ธ | S. Liu, T. Chen, X. Chen, L. Shen, D. C. Mocanu, Z. Wang, and M. Pechenizkiy, โThe unreasonable effectiveness of random pruning: Return of the most naive baseline for sparse training,โ in ICLR, 2022. | |
2025/04/16 | Pruning During Training: Sparsity Regularization based Methods | ๋ฐ์๋ฆฌ | ์จ๋ผ์ธ | W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li, โLearning structured sparsity in deep neural networks,โ in NIPS, 2016. | |
2025/04/23 | Pruning During Training: Dynamic Sparse Training based Methods | ๊ตฌ์น์ฐ | ์คํ๋ผ์ธ | U. Evci, T. Gale, J. Menick, P. S. Castro, and E. Elsen, โRigging the lottery: Making all tickets winners,โ in ICML, 2020. | |
2025/04/30 | Zero-shot Pruning | ์ ํ์ฐ | ์จ๋ผ์ธ | ์ ์ ์ค | Magical Week |
2025/05/07 | Pruning During Training: Score-based Methods | ์ตํด์ธ | ์จ๋ผ์ธ | Y. He, P. Liu, Z. Wang, Z. Hu, and Y. Yang, โFilter pruning via geometric median for deep convolutional neural networks acceleration,โ in CVPR, 2019, pp. 4340โ4349. | |
2025/05/14 | Pruning During Training: Differentiable Pruning based methods | ์ ์ง์ฐ | ์จ๋ผ์ธ | X. Ning, T. Zhao, W. Li, P. Lei, Y. Wang, and H. Yang, โDSA: More efficient budgeted pruning via differentiable sparsity allocation,โ in ECCV, 2020, pp. 592โ607. | Pseudo Con |
2025/05/21 | Pruning After Training: LTH and its Variants | ์ ํ์ฐ | ์จ๋ผ์ธ | ์ ์ ์ค | |
2025/05/28 | Pruning After Training: Other score-based Methods | ๊น๋ฏผ์ฑ | ์คํ๋ผ์ธ | ์ ์ ์ค | |
2025/06/04 | Pruning After Training: Sparsity Regularization based Methods | ์ต์์ | ์จ๋ผ์ธ | ์ ์ ์ค | |
2025/06/11 | Pruning After Training: Pruning in Early Training | ์๋ฌธ๊ธฐ | ์จ๋ผ์ธ | ์ ์ ์ค | |
2025/06/18 | Pruning After Training: Post-Training Pruning | ๋ฐ์ ์ | ์จ๋ผ์ธ | ์ ์ ์ค | |
2025/06/25 | Run-time Pruning | ์ ํ์ฐ | ์คํ๋ผ์ธ | ์ ์ ์ค |
๋ ์ง | ๋ด์ฉ | ์งํ๋ฐฉ์ | ๋น๊ณ |
---|---|---|---|
2025/04/01 | OT ๋ฐ ๊ณํ ์๋ฆฝ | ์จ๋ผ์ธ | |
2025/04/15 | Object Detection Model ์ ์ | ์จ๋ผ์ธ | |
2025/04/29 | [PDT] Sparsity Regularization based Method ๊ตฌํ ๋ฐ ํ ์คํธ | ์จ๋ผ์ธ | Magical Week |
2025/05/06 | ASP ๊ธฐ๋ฐ ๋ชจ๋ธ ํ์ต | ์จ๋ผ์ธ | |
2025/05/13 | TensorRT ๋ณํ ๋ฐ HW ๋ด ์ฑ๋ฅ ๋น๊ต | ์คํ๋ผ์ธ | PseudoCon |
2025/05/20 | [PDT] Sparse Training based Methods ๊ตฌํ ๋ฐ ํ ์คํธ | ์จ๋ผ์ธ | |
2025/05/27 | [PDT] Score-based Methods ๊ตฌํ ๋ฐ ํ ์คํธ | ์จ๋ผ์ธ | |
2025/06/03 | [PDT] Differentiable Pruning based methods ๊ตฌํ ๋ฐ ํ ์คํธ | ์จ๋ผ์ธ | |
2025/06/10 | [PDT] ๊ตฌํ๋ ๋ชจ๋ธ๋ค TensorRT ๋ณํ ๋ฐ HW ์ฑ๋ฅ ๋น๊ต | ์คํ๋ผ์ธ | |
2025/06/17 | [PAT] LTH and its Variants ๊ตฌํ ๋ฐ ํ ์คํธ | ์จ๋ผ์ธ | |
2025/06/24 | [PAT] Pruning in Early Training ๊ตฌํ ๋ฐ ํ ์คํธ | ์จ๋ผ์ธ | |
2025/07/01 | [PAT] Post-Training Pruning ๊ตฌํ ๋ฐ ํ ์คํธ | ์จ๋ผ์ธ | |
2025/07/08 | Run-time Pruning ๊ตฌํ ๋ฐ ํ ์คํธ | ์จ๋ผ์ธ |
๋งค์ฃผ ์คํฐ๋ ์งํ ๋ฐฉ์์ ๋ค์๊ณผ ๊ฐ์ต๋๋ค.
- ๊ทผํฉ ์ด์ผ๊ธฐ (20 ~ 30๋ถ ์์)
- ๋ฐํ์๋ฅผ ์ ์ธํ ์ฐธ์ฌ์๋ค์ด ์ค๋นํ On-Device AI ๊ด๋ จ๋ ์ด์๋ค์ ๊ณต์ ํ๋ค. (20 ~ 40๋ถ ์์)
- ๋ฐํ์๋ ์ค๋นํ ๋ ผ๋ฌธ ๋ฆฌ๋ทฐ๋ฅผ ๋ฐํํ๋ค. (30๋ถ ~ 1์๊ฐ ์์)
์ด์ ๋ฐ๋ผ ๋ค์ ๋ด์ฉ๋ค์ ์ค๋นํ์๋ฉด ๋ฉ๋๋ค
๊ณตํต์ฌํญ
- ํด๋น ์ฃผ์ฐจ ๋ ผ๋ฌธ์ ์ฝ๋๋ค.
๋ฐํ์
- ํด๋น ์ฃผ์ฐจ ๋ ผ๋ฌธ์ ๋ํ ๋ฐํ ์ค๋น๋ฅผ ํ๋ค.
์ฐธ์ฌ์
- On-Device AI์ ๊ด๋ จ๋ ๊ธฐ์ ๋ค(TensorRT, LiteRT, ONNX ๋ฑ)์ ํธ๋ ๋๋ ์ด์๋ฅผ ์ค๋นํ๋ค.
์ธ๋ถ ๋ ผ๋ฌธ๋ค์ ์ฃผ์ฐจ๋ณ ํ๋ ๋ด ์ฐธ๊ณ ์๋ฃ ์ฐธ๊ณ
์ฐธ๊ณ ๋ฌธํ
์งํ ์ ๋ณด
- ์๊ฐ: ๋งค์ฃผ ์์์ผ ์คํ 8์
- ์ฅ์: ์จ๋ผ์ธ / ์คํ๋ผ์ธ(๊ฐ๋จ์ญ)
์ฐธ์ฌ ์กฐ๊ฑด
- On-Device AI(๊ฒฝ๋ํ, ์ต์ ํ ๋ฑ)์ ๊ด์ฌ ์์ผ์ ๋ถ
- 4๊ฐ์ ๋์ ๊พธ์คํ ์ฐธ์ฌํ์ค ์ ์๋ ๋ถ
- ๋ฅ๋ฌ๋ ๊ธฐ์ด ์ง์ ๋ณด์ ํ์ ๋ถ
- ๋ ผ๋ฌธ์ ์ฝ๊ณ ๋ฆฌ๋ทฐํ์ค ์ ์๋ ๋ถ
ํ์์ผ๋ก ์ฐธ์ฌํ์๋ ค๋ฉด ๋ฌ๋ ๋ชจ์ง ๊ธฐ๊ฐ์ ์ ์ฒญํด์ฃผ์ธ์.
- ๋งํฌ (์ค๋น์ค)
๋๊ตฌ๋ ์ฒญ๊ฐ์ ํตํด ๋ชจ์์ ์ฐธ์ฌํ์ค ์ ์์ต๋๋ค.
- ํน๋ณํ ์ ์ฒญ ์์ด ์ ๊ธฐ ๋ชจ์ ์๊ฐ์ ๋ง์ถ์ด ๋์ค์ฝ๋ #Room-GH ์ฑ๋๋ก ์ ์ฅ
- Magical Week ์ค ํ์ฌ์ ์ฐธ๊ฐ
- Pseudo Lab ํ์ฌ์์ ๋ง๋๊ธฐ
Pseudo-Lab is a non-profit organization focused on advancing machine learning and AI technologies. Our core values of Sharing, Motivation, and Collaborative Joy drive us to create impactful open-source projects. With over 5k+ researchers, we are committed to advancing machine learning and AI technologies.
This project is licensed under the MIT License.