Тёмный

Relative Position Bias (+ PyTorch Implementation) 

Soroush Mehraban
Подписаться 3,1 тыс.
Просмотров 3,4 тыс.
50% 1

In this video, I explain why position embedding is required in vision transformers, what's the limitation of using absolute position embedding, and how relative position bias can improve that.
Table of Content:
00:00 Permutation Equivariance
01:12 Absolute Position Embedding
02:42 Limitation of absolute positions
03:56 Relative Position Bias intuition
07:57 Relative Position Bias in theory
12:53 PyTorch Implementation
Icon made by Freepik from flaticon.com

Опубликовано:

 

28 июл 2024

Поделиться:

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист
Посмотреть позже
Комментарии : 18   
@laspinetta2954
@laspinetta2954 13 дней назад
This part of the SwinTransformer paper is the least understood and took a long time, finally I understood it clearly thanks to this lecture. I would really appreciate it if you could find these points of many papers in the future and explain them easily! Super Thanks!
@soroushmehraban
@soroushmehraban 13 дней назад
@@laspinetta2954 I have to understand them first lol. I like to focus on them and make cool videos but recently I got so busy unfortunately
@proterotype
@proterotype 2 дня назад
Yeah I’m with this guy
@alighandour1057
@alighandour1057 7 месяцев назад
That's the best video explaining Relative Position Bias in detail on the internet, i hope your channel grows you deserve it!
@soroushmehraban
@soroushmehraban 7 месяцев назад
Thank you for the kind words Ali 🙂
@alihadimoghadam8931
@alihadimoghadam8931 Год назад
nice job pal ❤
@muttal-eg5pg
@muttal-eg5pg 9 месяцев назад
thank you
@qiliux
@qiliux 5 месяцев назад
Thanks for you amazing explanation. Can you take a look at convolutional relative positional encoding?
@65Jabulani
@65Jabulani Год назад
Hi, can you also explain the continuous relative position bias used in Swin V2?
@soroushmehraban
@soroushmehraban Год назад
Yeah sure! I will record a video for Swin V2 and explain that in that video.
@65Jabulani
@65Jabulani Год назад
@@soroushmehraban Thank you so much!! Your explanations are crystal clear. Very helpful
@soroushmehraban
@soroushmehraban Год назад
@@65Jabulani Glad you liked it
@arijaa.9315
@arijaa.9315 11 месяцев назад
Hi ! thanks for crstal explanation could you please provide a link to the code explained in the video?
@soroushmehraban
@soroushmehraban 11 месяцев назад
Here is the code: github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py#L101
@shilashm5691
@shilashm5691 Год назад
Transformer is Permutation Invariance, thats the reason of using position encodings
@muttal-eg5pg
@muttal-eg5pg 9 месяцев назад
permutation equivairant
@George70220
@George70220 10 месяцев назад
The vectors are added not concatenated, no? This is a very important fact.
@soroushmehraban
@soroushmehraban 10 месяцев назад
They're usually added to the input. Some works might concatenate it though. Both should make the model to consider the token orders.
Далее
Masked Autoencoders (MAE) Paper Explained
15:20
Просмотров 2,7 тыс.
CNN Receptive Field | Deep Learning Animated
10:28
Просмотров 2 тыс.
Swin Transformer V2 - Paper explained
21:32
Просмотров 2,9 тыс.
The Boundary of Computation
12:59
Просмотров 985 тыс.
Swin Transformer
9:23
Просмотров 1,4 тыс.
Diffusion models from scratch in PyTorch
30:54
Просмотров 239 тыс.