Comments (6)
tahmid0007
commented on July 3, 2024
Hi there,
Thanks for your interest in the work. First of all, I have to cross-check the differences but be assured that the performance does remain almost same with 1, 8 or even more than 8 attention heads. Same thing happens with the number of layers as well, doesn't really impact the ultimate outcome. If you explore more, you will find that these hyperparameters are taken casually most of the times. Nevertheless, I will update relevant parts if deemed necessary.
Thanks
from visualtransformers.
zhenglongjiepheonix
commented on July 3, 2024
I see, thanks for reply
from visualtransformers.
zhenglongjiepheonix
commented on July 3, 2024
by the way, do you know the difference between feature Map channel and visual token channel, I mean, they are both denoted by C in the paper and they seems to be the same according to the matrix multiplication, thanks for reply!
…------------------ 原始邮件 ------------------
发件人: "tahmid0007/VisualTransformers" <[email protected]>;
发送时间: 2021年1月12日(星期二) 上午7:27
收件人: "tahmid0007/VisualTransformers"<[email protected]>;
抄送: "Phoenix"<[email protected]>;"Author"<[email protected]>;
主题: Re: [tahmid0007/VisualTransformers] some questions (#6)
Hi there,
Thanks for your interest in the work. First of all, I have to cross-check the differences but be assured that the performance does remain almost same with 1, 8 or even more than 8 attention heads. Same thing happens with the number of layers as well, doesn't really impact the ultimate outcome. If you explore more, you will find that these hyperparameters are taken casually most of the times. Nevertheless, I will update relevant parts if deemed necessary.
Thanks
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
from visualtransformers.
tahmid0007
commented on July 3, 2024
If you look closely, you will see that they referred the feature map dimension as HWxC, here C is simply the number of filters. but for tokens, they refer it as LxC, where C is the width of each token. You can think of it like, each token has 1 participant from each of the C feature maps and each token has C length. I hope that makes your concept clear.
from visualtransformers.
zhenglongjiepheonix
commented on July 3, 2024
Yeah I mean. according to formula T = softmax_{HW}(XW_A)^T X , where X \in R^{HW,C}, here C is feature map dimension, according to the matrix multiplication rule, T must has the same last dimension C, and T here should be tokens, which means C in tokens and C in X, are the same?
…------------------ 原始邮件 ------------------
发件人: "tahmid0007/VisualTransformers" <[email protected]>;
发送时间: 2021年1月12日(星期二) 下午2:35
收件人: "tahmid0007/VisualTransformers"<[email protected]>;
抄送: "Phoenix"<[email protected]>;"Author"<[email protected]>;
主题: Re: [tahmid0007/VisualTransformers] some questions (#6)
If you look closely, you will see that they referred the feature map dimension as HWxC, here C is simply the number of filters. but for tokens, they refer it as LxC, where C is the width of each token. You can think of it like, each token has 1 participant from each of the C feature maps and has each token has C length. I hope that makes your concept clear.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
from visualtransformers.
zhenglongjiepheonix
commented on July 3, 2024
Hi again, I have contacted the author and he says I can use an additional matrix multiplication to change the dimension of tokens
…------------------ 原始邮件 ------------------
发件人: "tahmid0007/VisualTransformers" <[email protected]>;
发送时间: 2021年1月12日(星期二) 下午2:35
收件人: "tahmid0007/VisualTransformers"<[email protected]>;
抄送: "Phoenix"<[email protected]>;"Author"<[email protected]>;
主题: Re: [tahmid0007/VisualTransformers] some questions (#6)
If you look closely, you will see that they referred the feature map dimension as HWxC, here C is simply the number of filters. but for tokens, they refer it as LxC, where C is the width of each token. You can think of it like, each token has 1 participant from each of the C feature maps and has each token has C length. I hope that makes your concept clear.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
from visualtransformers.
Related Issues (12)
- Static Tokenization HOT 3
- Hi, regarding the nn1 of ViTResNet: HOT 3
- Semantic Segmentation HOT 1
- Code for Visual Transformer or Vision Transformer? HOT 3
- No Documentation! HOT 3
- Question on token HOT 3
- Did u training a smeantic segmentation example using this model?aa
- Classification tokens HOT 2
- Performances
- Mask shape is not correct HOT 1
- Only 'BasicBlock',not have 'Bottleneck'
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from visualtransformers.