Vision Permutator: A Permutable MLP-Like Architecture for Visual Recognition
— AK (@ak92501) June 24, 2021
pdf: https://t.co/2Kb6zoFwR1
github: https://t.co/Ahe17BEDwf
achieves 81.5% top-1 accuracy on ImageNet without extra large-scale training data (e.g., ImageNet-22k) using only 25M learnable parameters pic.twitter.com/KO1Kx7OnSi
How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers
— AK (@ak92501) June 21, 2021
pdf: https://t.co/6isx98PS1M
train ViT models of various sizes on the
ImageNet-21k dataset, either match or outperform their counterparts trained on larger, but not publicly available JFT-300M pic.twitter.com/Zoso6tUPPA
Skip-Sideways is an efficient approximation of backprop for training video networks. It allows training on potentially infinite video sequences, being independent of the video length: https://t.co/OXG2AgKvlb (1/) pic.twitter.com/JxnJk9FzIN
— DeepMind (@DeepMind) June 18, 2021
🤗Transformers v4.7.0 was just released with 🖼️DETR by @facebookai!
— Hugging Face (@huggingface) June 17, 2021
DETR is an Object Detection model that can take models from timm by @wightmanr as a backbone.
Contributed by @NielsRogge, try it out: https://t.co/0AOf3P7QaC
v4.7.0 launches with support for PyTorch v1.9.0! pic.twitter.com/1LAYy4cN2W
Watching Too Much Television is Good: Self-Supervised Audio-Visual Representation Learning from Movies and TV Shows
— AK (@ak92501) June 17, 2021
pdf: https://t.co/hBN7Um0RMm
abs: https://t.co/cRXjYZzUhy pic.twitter.com/t56CsoEYPx
This is stellar work in collaboration with folks in Zurich. Do neural networks really make overconfident predictions in general? For new architectures, this is not true. It may not even be true consistently for CNNs! @FrancesAnnHubis @XiaohuaZhai @neilhoulsby @MarioLucic_ https://t.co/zkB1zzmzkp
— Dustin Tran (@dustinvtran) June 16, 2021
Keep CALM and Improve Visual Feature Attribution
— AK (@ak92501) June 16, 2021
pdf: https://t.co/I90yEPbjse
abs: https://t.co/JKIlsnNnGn
github: https://t.co/MP1F8LXvY0
identifies discriminative attributes for image classifiers more accurately than CAM and other visual attribution baselines pic.twitter.com/50DBMoVYai
BEIT: BERT Pre-Training of Image Transformers
— AK (@ak92501) June 16, 2021
pdf: https://t.co/WiFZIiErLt
abs: https://t.co/Ld2067ltiV
large-size BEIT obtains 86.3% only using ImageNet-1K, even outperforming ViT-L with supervised
pre-training on ImageNet-22K (85.2%) pic.twitter.com/abMaWZ1aZ8
Styleformer: Transformer based Generative Adversarial Networks with Style Vector
— AK (@ak92501) June 15, 2021
pdf: https://t.co/jNVLty3unL
abs: https://t.co/SEK0ko63E7
github: https://t.co/hQanKidsZ8
outperforms GAN-based generative models, including StyleGAN2-ADA with fewer parameters on CIFAR-10 pic.twitter.com/bs3JmTJtdz
MlTr: Multi-label Classification with Transformer
— AK (@ak92501) June 14, 2021
pdf: https://t.co/wqvd89AtJq
abs: https://t.co/H2n64N5OGa pic.twitter.com/yXFM5caLMK
SimSwap: An Efficient Framework For High Fidelity
— AK (@ak92501) June 14, 2021
Face Swapping
pdf: https://t.co/l2aWTrM1CP
abs: https://t.co/ZSuDnRLUuF
github: https://t.co/deYKr8rhLY pic.twitter.com/cBuaXySkd9
Today, we’re introducing TextStyleBrush, the first self-supervised AI model that replaces text in existing images of both scenes and handwriting — in one shot — using just a single example word: https://t.co/0QfLraAQvV pic.twitter.com/FNDJxNC20S
— Facebook AI (@facebookai) June 11, 2021
