๋ฐ˜์‘ํ˜•

๐ŸŒŒ Deep Learning 50

[๋”ฅ๋Ÿฌ๋‹ ๋…ผ๋ฌธ๋ฆฌ๋ทฐ] PointMLP - Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework (ICLR 2022)

ICLR 2022์—์„œ ๋ฐœํ‘œ๋œ Pointcloud analysis๋ฅผ ์œ„ํ•œ ๋ชจ๋ธ์ด๋‹ค. ๊ธฐ์กด Pointcloud analysis ๋ชจ๋ธ๋“ค์€ ๋ณต์žกํ•œ local geometric extractor๋“ค์„ ์ด์šฉํ•ด ์™”๋Š”๋ฐ, ์ด๋“ค๊ณผ ๋‹ฌ๋ฆฌ ๋งค์šฐ ๊ฐ„๋‹จํ•œ ๊ตฌ์กฐ๋กœ SOTA๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์˜€๋‹ค. Paper, Code 1. Introduction ๊ธฐ์กด pointcloud ๋ชจ๋ธ๋“ค์€ 3D geometric์„ ํฌ์ฐฉํ•˜๊ธฐ ์œ„ํ•ด convolution, graph, attention ๋“ฑ ๋งค์šฐ ๋ณต์žกํ•œ local geometric extractor๋“ค์„ ์ด์šฉํ–ˆ๋Š”๋ฐ, ์ด๋“ค์€ computational overhead๊ฐ€ ๋„ˆ๋ฌด ํฌ๋‹ค๋Š” ๋‹จ์ ์ด ์žˆ๋‹ค. ์ €์ž๋“ค์€ pointcloud analysis์˜ ํ•ต์‹ฌ์€ detailed local geometrical i..

[StyleGAN ์‹œ๋ฆฌ์ฆˆ] ProGAN/PGGAN, StyleGAN, StyleGAN2

ProGAN๋ถ€ํ„ฐ StyleGAN2๊นŒ์ง€, style transfer์—์„œ ๊ฐ€์žฅ ์œ ๋ช…ํ•œ ๋ชจ๋ธ์ธ StyleGAN์˜ ๋ณ€์ฒœ์‚ฌ์™€ ๊ฐ ๋ชจ๋ธ์˜ ํŠน์ง•์„ ๊ฐ„๋‹จํžˆ ์ •๋ฆฌํ•ด ๋ณด๊ณ ์ž ํ•œ๋‹ค. 1. ProGAN/PGGAN (ICLR 2018) Paper: Progressive Growing of GANs for Improved Quality, Stability, and Variation (link) GAN์„ ์ด์šฉํ•ด ๊ณ ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์‰ฝ์ง€ ์•Š๋‹ค. ๋”ฐ๋ผ์„œ latent vector์—์„œ ํ•œ๋ฒˆ์— ๊ณ ํ•ด์ƒ๋„์˜ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ๋ณด๋‹ค๋Š”, ๋‚ฎ์€ ํ•ด์ƒ๋„์˜ ์ด๋ฏธ์ง€(4x4)๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ๋ถ€ํ„ฐ ํ•™์Šตํ•ด์„œ ์ ์ง„์ ์œผ๋กœ(progressive) ๋ ˆ์ด์–ด๋ฅผ ์ถ”๊ฐ€ํ•˜๋ฉฐ ๊ณ ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€(1024x1024)๋ฅผ ๋งŒ๋“œ๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•œ๋‹ค. ๋ ˆ์ด์–ด๋ฅผ ์ถ”๊ฐ€ํ•  ๋•Œ๋Š” fade..

[๋”ฅ๋Ÿฌ๋‹ ๋…ผ๋ฌธ๋ฆฌ๋ทฐ] SimSiam: Exploring Simple Siamese Representation Learning (CVPR 2021)

CVPR 2021์—์„œ ๋ฐœํ‘œ๋œ self-supervised learning ๋…ผ๋ฌธ. ๊ธฐ์กด์˜ ๋‹ค๋ฅธ self-supervised learning ๋ฐฉ๋ฒ•์ธ SimCLR, SwAV, BYOL๊ณผ์˜ ๋น„๊ต๋ฅผ ํ†ตํ•ด ์ œ์•ˆ๋œ ๋ฐฉ๋ฒ•์„ ์„ค๋ช…ํ•˜๊ณ  ์žˆ๋‹ค. ๊ฐœ์ธ์ ์œผ๋กœ ๋ฌธ์žฅ์ด๋‚˜ ์ „์ฒด์ ์ธ ๊ตฌ์กฐ๊ฐ€ ๋งค์šฐ ์ดํ•ดํ•˜๊ธฐ ์‰ฝ๊ฒŒ ์ž˜ ์“ฐ์—ฌ์ ธ ์žˆ๋‹ค๊ณ  ๋Š๊ผˆ๋‹ค. ๋…ผ๋ฌธ ๋งํฌ Introduction Self-supervsied learning์—์„œ๋Š” ๋ณดํ†ต Siamese network ๊ตฌ์กฐ๋ฅผ ๋งŽ์ด ์ด์šฉํ•œ๋‹ค. ์ด๋Š” weight๋ฅผ ์„œ๋กœ ๊ณต์œ ํ•˜๋Š” neural network๋ฅผ ์˜๋ฏธํ•˜๋Š”๋ฐ, ์ด๋“ค์€ ๊ฐ entity๋ฅผ ๋น„๊ตํ•˜๋Š” ๋ฐ์— ์œ ์šฉํ•˜๊ฒŒ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Siamese network๋Š” output์ด ํ•˜๋‚˜์˜ constant๋กœ ์ˆ˜๋ ดํ•˜๋Š” collapsing์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋‹ค..

[PyTorch Implementation] PointNet ์„ค๋ช…๊ณผ ์ฝ”๋“œ

PointCloud ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•œ ๋Œ€ํ‘œ์ ์ธ ๋ชจ๋ธ์ธ PointNet์˜ ๊ตฌ์กฐ์™€ PyTorch๋กœ ๊ตฌํ˜„ํ•œ ์ฝ”๋“œ์ด๋‹ค. PointNet์€ Feature extraction ํ›„ classification / segmentation์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋ณธ ๊ธ€์—์„œ๋Š” classification์„ ์œ„ํ•œ ๋„คํŠธ์›Œํฌ๋งŒ ์†Œ๊ฐœํ•œ๋‹ค. ์ฝ”๋“œ๋Š” ๊ฐ€์žฅ star ์ˆ˜๊ฐ€ ๋งŽ์€ PyTorch implementation์ธ ์•„๋ž˜ Github repo๋ฅผ ์ฐธ๊ณ ํ•˜์—ฌ ์ผ๋ถ€ ์ˆ˜์ •ํ–ˆ๋‹ค. https://github.com/fxia22/pointnet.pytorch/tree/f0c2430b0b1529e3f76fb5d6cd6ca14be763d975 64๋กœ ๋Š˜๋ ค์ค€๋‹ค. 64์ฐจ์›์˜ shared mlp์— ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ T-Net๊ณผ matrix multiplication์„ ..

[๋”ฅ๋Ÿฌ๋‹ ๋…ผ๋ฌธ๋ฆฌ๋ทฐ + ์ฝ”๋“œ] What uncertainties do we need in Bayesian deep learning for computer vision? (NeurIPS 2017)

๋…ผ๋ฌธ: https://arxiv.org/pdf/1703.04977.pdf Epistemic uncertainty์™€ aleatoric uncertainty๋ฅผ ๋™์‹œ์— ์ธก์ •ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ฃผ๋Š” ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•œ ๋…ผ๋ฌธ์ด๋‹ค. ๋…ผ๋ฌธ์„ ๊ฐ„๋‹จํžˆ ์ •๋ฆฌํ•˜๊ณ  PyTorch ์ฝ”๋“œ๋ฅผ ํ•จ๊ป˜ ์†Œ๊ฐœํ•˜๊ณ ์ž ํ•œ๋‹ค. Uncertainty์˜ ์ข…๋ฅ˜ 1. Epistemic uncertainty (=Model uncertainty) ๋ชจ๋ธ๊ตฌ์กฐ๋‚˜ ํ•™์Šต๊ณผ์ •์—์„œ ๋ฐœ์ƒํ•˜๋Š” uncertainty์ด๋‹ค. ๋ชจ๋ธ์ด ์ถฉ๋ถ„ํžˆ ํ•™์Šต๋˜์ง€ ์•Š์•˜์„ ์ˆ˜๋„ ์žˆ๊ณ , ์ „์ฒด ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ๋‹ค ํ•™์Šตํ•˜์ง€ ๋ชปํ–ˆ์„ ์ˆ˜๋„ ์žˆ๊ณ , ๋ชจ๋ธ์˜ ๊ตฌ์กฐ๊ฐ€ ์ง€๋‚˜์น˜๊ฒŒ ๋‹จ์ˆœํ•˜๊ฑฐ๋‚˜ ๋ณต์žกํ•  ์ˆ˜๋„ ์žˆ๋‹ค. Epistemic uncertainty๋Š” ๋ฐ์ดํ„ฐ์…‹ ๋ณด๊ฐ•, ๋ชจ๋ธ ๊ตฌ์กฐ ์ˆ˜์ •, ํ•™์Šต ๋ฐฉ๋ฒ• ๋ณ€๊ฒฝ ๋“ฑ์˜ ๋ฐฉ๋ฒ•์œผ๋กœ ์ค„์ผ ์ˆ˜ ..

๋”ฅ๋Ÿฌ๋‹ ๋…ผ๋ฌธ ์ €์ž๋“ค์˜ paper talk ์•„์นด์ด๋ธŒ ์‚ฌ์ดํŠธ Papertalk

https://papertalk.org/index Papertalk - the platform for scientific paper presentations You have to login to use this feature. papertalk.org ๋”ฅ๋Ÿฌ๋‹ ๋…ผ๋ฌธ ์ €์ž๋“ค์˜ paper talk์„ ๋ชจ์•„ ๋‘” ์‚ฌ์ดํŠธ์ด๋‹ค. ์ตœ์‹  ๋…ผ๋ฌธ, Journal/Conference ๋ณ„, ์ธ๊ธฐ์ˆœ, ํ‚ค์›Œ๋“œ ๋ณ„๋กœ ๋…ผ๋ฌธ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๊ณ , ๋””์ž์ธ์ด ๋งค์šฐ ๊น”๋”ํ•˜๊ฒŒ ์ •๋ฆฌ๋˜์–ด ์žˆ์–ด์„œ ๋…ผ๋ฌธ์˜ overview๊ฐ€ ํ•„์š”ํ•  ๋•Œ ์œ ์šฉํ•˜๊ฒŒ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ ๊ฐ™๋‹ค. ๐Ÿ˜ฎ๐Ÿ‘ ๊ธฐ๋ณธ์ ์œผ๋กœ๋Š” ํ•™ํšŒ์˜ ๋ฐœํ‘œ์˜์ƒ์ด ์˜ฌ๋ผ๊ฐ€์žˆ๊ณ , ์ €์ž๊ฐ€ ์ง์ ‘ ์š”์ฒญํ•˜๋ฉด ๋‹ค๋ฅธ ์˜์ƒ์œผ๋กœ ๋ณ€๊ฒฝํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•œ๋‹ค. ๋‹ค๋งŒ 2021๋…„๊นŒ์ง€์˜ paper talk์€ ์•„์นด์ด๋ธŒ๊ฐ€ ๋˜์–ด ์žˆ๋Š”๋ฐ, 2022๋…„..

[๋”ฅ๋Ÿฌ๋‹ ๋…ผ๋ฌธ๋ฆฌ๋ทฐ] Momentum Contrast for Unsupervised Visual Representation Learning (MoCo) (CVPR 2020)

SimCLR์™€ ํ•จ๊ป˜ ๊ฐ€์žฅ ์œ ๋ช…ํ•œ contrastive learning-based self-supervised learning ๋…ผ๋ฌธ์ด๋‹ค. CVPR 2020์—์„œ ๋ฐœํ‘œ๋œ ๋…ผ๋ฌธ์ด๊ณ , Kaiming He๊ฐ€ ์ €์ž๋กœ ์ฐธ์—ฌํ•˜์˜€๋‹ค. Paper, Code GitHub - facebookresearch/moco: PyTorch implementation of MoCo: https://arxiv.org/abs/1911.05722 PyTorch implementation of MoCo: https://arxiv.org/abs/1911.05722 - GitHub - facebookresearch/moco: PyTorch implementation of MoCo: https://arxiv.org/abs/1911.05722 gith..

[๋”ฅ๋Ÿฌ๋‹ ๋…ผ๋ฌธ๋ฆฌ๋ทฐ + ์ฝ”๋“œ] Uncertainty-Driven Loss for Single Image Super-Resolution (NeurIPS 2021)

NeurIPS 2021์—์„œ ๋ฐœํ‘œ๋œ ๋…ผ๋ฌธ์œผ๋กœ, single image super-resolution (SISR)์„ ์œ„ํ•ด uncertainty๋ฅผ ์ด์šฉํ•œ loss๋ฅผ ์ œ์•ˆํ–ˆ๋‹ค. ๋งจ ์•„๋ž˜์— PyTorch๋กœ ๊ตฌํ˜„๋œ ์ฝ”๋“œ๋ฅผ ์ •๋ฆฌํ•ด ๋†“์•˜๋‹ค. ๋…ผ๋ฌธ ๋งํฌ: https://papers.nips.cc/paper/2021/file/88a199611ac2b85bd3f76e8ee7e55650-Paper.pdf Supplementary: https://papers.nips.cc/paper/2021/file/88a199611ac2b85bd3f76e8ee7e55650-Supplemental.pdf Homepage: https://see.xidian.edu.cn/faculty/wsdong/Projects/UDL-SR.htm https:/..

[PyTorch Implementation] ResNet-B, ResNet-C, ResNet-D, ResNet Tweaks

Bag of Tricks for Image Classification with Convolutional Neural Networks (He et al., CVPR 2019)์—์„œ ์†Œ๊ฐœ๋œ ResNet์˜ ๋ณ€ํ˜• ๋ชจ๋ธ๋“ค(tweaks)์— ๋Œ€ํ•œ PyTorch ์ฝ”๋“œ์ด๋‹ค. ๋…ผ๋ฌธ์—์„œ๋Š” ๋‘ ๊ฐœ์˜ ์œ ๋ช…ํ•œ ๋ณ€ํ˜• ๋ชจ๋ธ์ธ ResNet-B, C๋ฅผ ์†Œ๊ฐœํ•˜๊ณ , ์ƒˆ๋กœ์šด ๊ตฌ์กฐ์ธ ResNet-D๋ฅผ ์ œ์•ˆํ•˜๊ณ  ์žˆ๋‹ค. ResNet ๊ธฐ๋ณธ์ ์ธ ResNet์˜ ๊ตฌ์กฐ๋Š” Input stem๊ณผ 4๊ฐœ์˜ stage, ๊ทธ๋ฆฌ๊ณ  ๋งˆ์ง€๋ง‰ output layer๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋‹ค. Input stem : 7x7 conv์™€ maxpool๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ์œผ๋ฉฐ, channel dim์„ 64๋กœ ๋Š˜๋ฆฌ๊ณ  input size๋ฅผ 4๋ฐฐ ์ค„์ธ๋‹ค. Stage : ํ•œ ๊ฐœ์˜ downsampling..

[PyTorch Implementation] CBAM: Convolutional Block Attention Module ์„ค๋ช… + ์ฝ”๋“œ

CBAM: Convolutional Block Attention Module์€ ECCV 2018์—์„œ ๋ฐœํ‘œ๋œ channel&spatial attention module์ด๋‹ค. ์ฝ”๋“œ๋Š” ๊ณต์‹ github์„ ์ฐธ๊ณ ํ•˜์—ฌ ์กฐ๊ธˆ ์ˆ˜์ •ํ–ˆ๋‹ค. Paper: https://arxiv.org/pdf/1807.06521.pdf Code: https://github.com/Jongchan/attention-module/ Author's blog: https://blog.lunit.io/2018/08/30/bam-and-cbam-self-attention-modules-for-cnn/ BAM CBAM์€ BAM: Bottleneck Attention Module์˜ ํ›„์† ๋…ผ๋ฌธ์ด๋‹ค. BAM์€ ๋ชจ๋ธ์˜ bottleneck ๋ถ€๋ถ„์—์„œ attent..

๋ฐ˜์‘ํ˜•