Konto erstellenAnmelden

Frühere Suchanfragen

Keine früheren Suchanfragen

Suchoptionen

Nur verfügbar, wenn angemeldet.
sueden.social ist einer von vielen unabhängigen Mastodon-Servern, mit dem du dich im Fediverse beteiligen kannst.
Eine Community für alle, die sich dem Süden hingezogen fühlen. Wir können alles außer Hochdeutsch.

Verwaltet von:

Serverstatistik:

2 Tsd.
aktive Profile

sueden.social: ÜberStatusProfilverzeichnisDatenschutzerklärung

Mastodon: ÜberApp herunterladenTastenkombinationenQuellcode anzeigenv4.3.7

#DeepLearning

7 Beiträge6 Beteiligte0 Beiträge heute
Öffentlich
Hacker News

Vision Transformers Need Registers (extra tokens discarded after last layer)

arxiv.org/abs/2309.16588

arXiv logo
arXiv.orgVision Transformers Need RegistersTransformers have recently emerged as a powerful tool for learning visual representations. In this paper, we identify and characterize artifacts in feature maps of both supervised and self-supervised ViT networks. The artifacts correspond to high-norm tokens appearing during inference primarily in low-informative background areas of images, that are repurposed for internal computations. We propose a simple yet effective solution based on providing additional tokens to the input sequence of the Vision Transformer to fill that role. We show that this solution fixes that problem entirely for both supervised and self-supervised models, sets a new state of the art for self-supervised visual models on dense visual prediction tasks, enables object discovery methods with larger models, and most importantly leads to smoother feature maps and attention maps for downstream visual processing.
#HackerNews#VisionTransformers#Registers
Öffentlich
Hacker News

Three things everyone should know about Vision Transformers

arxiv.org/abs/2203.09795

arXiv logo
arXiv.orgThree things everyone should know about Vision TransformersAfter their initial success in natural language processing, transformer architectures have rapidly gained traction in computer vision, providing state-of-the-art results for tasks such as image classification, detection, segmentation, and video analysis. We offer three insights based on simple and easy to implement variants of vision transformers. (1) The residual layers of vision transformers, which are usually processed sequentially, can to some extent be processed efficiently in parallel without noticeably affecting the accuracy. (2) Fine-tuning the weights of the attention layers is sufficient to adapt vision transformers to a higher resolution and to other classification tasks. This saves compute, reduces the peak memory consumption at fine-tuning time, and allows sharing the majority of weights across tasks. (3) Adding MLP-based patch pre-processing layers improves Bert-like self-supervised training based on patch masking. We evaluate the impact of these design choices using the ImageNet-1k dataset, and confirm our findings on the ImageNet-v2 test set. Transfer performance is measured across six smaller datasets.
#HackerNews#VisionTransformers#AI
Öffentlich
Météo Blois

Bon ce soir pas bp de temps, alors on fait un modèle #deeplearning quick & dirty vite fait histoire d'avoir une base de code.
On part du géopotentiel Z500 (ERA5) et on tente de prédire l'échéance suivante avec qques couches de neurones. Alors évidemment le résultat est à la hauteur de l'effort 😅

Un exemple de géopotentiel global issu de ERA5 (donnée d'entrée)
Une prédiction faite avec le modèle... Bah c'est rien d'autre que du bruit pour l'instant. On va tenter de faire mieux étape par étape ^^
Öffentlich *
QGIS Gebruikers Nederland

🤖 Afgelopen vrijdag hebben wij bij #ITC
@utwente de toekomst van #AI #MachineLearning en #DeepLearning in #QGIS verkend. Inspirerende presentaties over plugins en tools. Waardevolle discussies en inzichten voor de toekomst van deze technologie in QGIS. Dank aan alle deelnemers en in het bijzonder onze internationale gasten Matthias Kuhn en @suricactus van @opengisch.

Dank aan Rosa Aguilar voor het organiseren en de steun van @eScienceCenter!

#QGISNL#OSGeoNL#OSGeo
EntdeckenLive-Feeds
Über