Latest AI news 2023/12/29 ์ตœ์‹  AI ๋‰ด์ŠคMobileVLM : A Fast, Reproducible and Strong Vision Language Assistant for Mobile Devices
MobileVLM : ๋ชจ๋ฐ”์ผ ๊ธฐ๊ธฐ๋ฅผ ์œ„ํ•œ ๋น ๋ฅด๊ณ , ์žฌํ˜„ ๊ฐ€๋Šฅํ•˜๋ฉฐ ๊ฐ•๋ ฅํ•œ ์‹œ๊ฐ ์–ธ์–ด ๋ณด์กฐ๊ธฐ

Meituan Inc

Zhejiang University, China

Dalian University of Technology, China


3์ค„์š”์•ฝ

  1. MobileVLM์€ ๋ชจ๋ฐ”์ผ ๊ธฐ๊ธฐ์—์„œ ์‹คํ–‰๋˜๋„๋ก ์„ค๊ณ„๋œ ๋‹ค์ค‘๋ชจ๋‹ฌ ๋น„์ „ ์–ธ์–ด ๋ชจ๋ธ๋กœ, 1.4B ๋ฐ 2.7B ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ทœ๋ชจ์˜ ์–ธ์–ด ๋ชจ๋ธ๋“ค๊ณผ CLIP ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋‹ค์ค‘๋ชจ๋‹ฌ ๋น„์ „ ๋ชจ๋ธ์„ ํ†ตํ•ฉํ•˜์˜€์Šต๋‹ˆ๋‹ค.
  2. ๋ชจ๋ธ์€ ์—ฌ๋Ÿฌ VLM ๋ฒค์น˜๋งˆํฌ์—์„œ ํ‰๊ฐ€๋˜์—ˆ์œผ๋ฉฐ, ํฐ ๋ชจ๋ธ๋“ค๊ณผ ๋น„๊ตํ•˜์—ฌ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์œผ๋ฉฐ, Qualcomm Snapdragon 888 CPU์™€ NVIDIA Jeston Orin GPU์—์„œ ๋†’์€ ์ถ”๋ก  ์†๋„๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.
  3. ์ด ์—ฐ๊ตฌ์˜ ์ฝ”๋“œ๋Š” GitHub์—์„œ ๊ณต๊ฐœ๋  ์˜ˆ์ •์ด๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ฐ”์ผ ๊ธฐ๊ธฐ๋ฅผ ์œ„ํ•œ ๊ฐ•๋ ฅํ•˜๊ณ  ํšจ์œจ์ ์ธ ๋น„์ „ ์–ธ์–ด ๋ณด์กฐ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Abstract

We present MobileVLM, a competent multimodal vision language model (MMVLM) targeted to run on mobile devices. It is an amalgamation of a myriad of architectural designs and techniques that are mobile-oriented, which comprises a set of language models at the scale of 1.4B and 2.7B parameters, trained from scratch, a multimodal vision model that is pre-trained in the CLIP fashion, cross-modality interaction via an efficient projector. We evaluate MobileVLM on several typical VLM benchmarks. Our models demonstrate on par performance compared with a few much larger models. More importantly, we measure the inference speed on both a Qualcomm Snapdragon 888 CPU and an NVIDIA Jeston Orin GPU, and we obtain state-of-the-art performance of 21.5 tokens and 65.3 tokens per second, respectively.


์šฐ๋ฆฌ๋Š” ๋ชจ๋ฐ”์ผ ๊ธฐ๊ธฐ์—์„œ ์‹คํ–‰๋˜๋„๋ก ์„ค๊ณ„๋œ ์œ ๋Šฅํ•œ ๋‹ค์ค‘๋ชจ๋‹ฌ ๋น„์ „ ์–ธ์–ด ๋ชจ๋ธ(MMVLM)์ธ MobileVLM์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ์€ 1.4B ๋ฐ 2.7B ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ทœ๋ชจ์˜ ์–ธ์–ด ๋ชจ๋ธ๋“ค๋กœ ๊ตฌ์„ฑ๋œ ๋ชจ๋ฐ”์ผ ์ค‘์‹ฌ์˜ ๋‹ค์–‘ํ•œ ๊ฑด์ถ• ์„ค๊ณ„ ๋ฐ ๊ธฐ์ˆ ๋“ค์˜ ๊ฒฐํ•ฉ์ฒด์ด๋ฉฐ, ์ด ๋ชจ๋ธ๋“ค์€ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จ๋˜์—ˆ๊ณ , CLIP ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋‹ค์ค‘๋ชจ๋‹ฌ ๋น„์ „ ๋ชจ๋ธ, ํšจ์œจ์ ์ธ ํ”„๋กœ์ ํ„ฐ๋ฅผ ํ†ตํ•œ ๊ต์ฐจ ๋ชจ๋‹ฌ ์ƒํ˜ธ์ž‘์šฉ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” MobileVLM์„ ์—ฌ๋Ÿฌ ์ „ํ˜•์ ์ธ VLM ๋ฒค์น˜๋งˆํฌ์—์„œ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ๋“ค์€ ๋ช‡๋ช‡ ํ›จ์”ฌ ํฐ ๋ชจ๋ธ๋“ค๊ณผ ๋น„๊ตํ•˜์—ฌ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋” ์ค‘์š”ํ•œ ๊ฒƒ์€, ์šฐ๋ฆฌ๋Š” Qualcomm Snapdragon 888 CPU์™€ NVIDIA Jeston Orin GPU์—์„œ ์ถ”๋ก  ์†๋„๋ฅผ ์ธก์ •ํ•˜๋ฉฐ, ๊ฐ๊ฐ ์ดˆ๋‹น 21.5 ํ† ํฐ ๋ฐ 65.3 ํ† ํฐ์˜ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.



24b0d121e09c28a8699fe8b115ef046b696a95342a


24b0d121e09c28a8699fe8b115ef046a786fe8cd


24b0d121e09c28a8699fe8b115ef046ecc4bcff0


24b0d121e09c28a8699fe8b115ef046543aefb3f




https://arxiv.org/pdf/2312.16886.pdf

https://github.com/Meituan-AutoML/MobileVLM

GitHub - Meituan-AutoML/MobileVLM: MobileVLM : A Fast, Reproducible and Strong Vision Language Assistant for Mobile DevicesMobileVLM : A Fast, Reproducible and Strong Vision Language Assistant for Mobile Devices - GitHub - Meituan-AutoML/MobileVLM: MobileVLM : A Fast, Reproducible and Strong Vision Language Assistant...github.com