Скачать или смотреть DINO in PyTorch

In this video, we implement the DINO scheme in PyTorch. It is a selfsupervised learning framework for images proposed in the paper "Emerging Properties in Self-Supervised Vision Transformers". The goal is to train a network that is able to represent images in a meaningful way without using human labels. Among other things, if applied to a Vision Transformer network one can inspect the attention weights and get surprisingly good results. Last but not least, this video also contains short tutorials on weight normalization and PyTorch buffers.

Paper: https://arxiv.org/abs/2104.14294
Official code: https://github.com/facebookresearch/dino
Imagenette dataset: https://github.com/fastai/imagenette
Vision Transformer code: https://github.com/rwightman/pytorch-...
Code from the video: https://github.com/jankrepl/mildlyove...


00:00 Intro + disclaimer
00:50 High-level explanation
02:38 Imagenette - our dataset
03:11 timm - our Vision Transformer
03:29 Augmentations implemented
08:01 Augmentations visualized
08:48 Head implementation
11:12 [IPython] Weight normalization
15:48 Head implementation continued
16:54 MultiCropWrapper implementation
18:48 [IPython] Buffers in PyTorch
21:13 Loss implementation
25:50 Gradient clipping implementation
26:24 KNN classifier implementation
28:39 TensorBoard embedding preparation
30:01 Training script
37:35 Launch training
38:04 Results: Supervised model
39:17 Results: Selfsupervised model
41:17 Results: Visualizing attentions
43:39 Outro

If you have any video suggestions or you just wanna chat feel free to join the discord server:   / discord  

Twitter:   / moverfitted  

Credits logo animation
Title: Conjungation · Author: Uncle Milk · Source:   / unclemilk   · License: https://creativecommons.org/licenses/... · Download (9MB): https://auboutdufil.com/?id=600