zijian-hu / simple Goto Github PK
View Code? Open in Web Editor NEWCode for the paper: "SimPLE: Similar Pseudo Label Exploitation for Semi-Supervised Classification"
Home Page: https://arxiv.org/abs/2103.16725
License: Apache License 2.0
Code for the paper: "SimPLE: Similar Pseudo Label Exploitation for Semi-Supervised Classification"
Home Page: https://arxiv.org/abs/2103.16725
License: Apache License 2.0
Hello, Zijian Hu
I have a question because the mini-imagenet file uploaded to github is no longer downloaded.
I would be really grateful if you could tell me how you configured train/test of 50000/10000 about mini-imagenet dataset.
If there is a file, I would be very grateful if you could give me a link.
I look forward to your kind response. Thank you once again for your valuable time.
Hi, thanks for your work!
In your paper, Tab.3, Res18 can achieve 49.39 on miniImagenet.
Here is my wandb log (the code is your repo):
The bottom one is the baseline by disabling PairLoss. Top two are for SimPLE on two different machines (same config). It can be seen that the accuracy curve is almost flat. I think it is hard to get additional 5+% in the following training steps. And I checked your code. You don't use any lr scheduler. So I think the acc. may not increase sharply in the following step.
Could you please share a config or a wandb log of resnet18-miniimagenet?
Best wishes.
Thanks a lot for your released codes and your excellent work. I have some questions about the reported results in your paper.
(1) Does all reported results in the paper obtained by the final EMA model instead of the trainable student model?
(2) Does the reported results are obtained by calculating the mean accuracy value of the latest several (e.g., 20) iterations/epochs during the training phase or just select the best accuracy of someone iteration/epoch, or any other protocols?
Dear Zijian,
Thank you for your contribution! I got a question when reading the pair loss computing part.
I got confused about whether the similarity should be computed between pseudo labels generated from weak augmented data and prediction of strong augmented data, or, pseudo labels generated from weak augmented data among different input pictures.
If the latter one is right, does line 70 means that: if a weak-augmented-image(A)-generated pseudo label with high confidence is similar enough to the pseudo label from another input image B, then the distance between pseudo label A and prediction of strong augmented image B should be minimized?
distance_ij = self.get_distance_loss(loss_input, targets_i, dim=1, reduction='none')
Thank you for your time. I would appreciate it if I would be replied.
Best,
Vivian
Hi, thanks for your excellent work.
When will you release your code?
Thanks
Dear zijianhu
Hello, I can't download the domainnet_real dataset you used, can you fix it to the latest version?
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. ๐๐๐
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.