albert0147 / nrc_sfda Goto Github PK
View Code? Open in Web Editor NEWCode for our NeurIPS 2021 paper 'Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation'
License: MIT License
Code for our NeurIPS 2021 paper 'Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation'
License: MIT License
Thank your for nice job!
I notice that you use dot-product to measure simlarity between prediction in paper while use KL_div in code.
code:
NRC_SFDA/office-home/train_tar.py
Lines 315 to 317 in 1c26160
First of all, thanks for your excellent work.
I have tried to reproduce the results of VISDA-C with your provided source model and codes, and I get a log as follows:
Task: TV, Iter:866/12990; Accuracy on target = 82.84%
T: 95.86 82.47 83.18 65.42 94.33 96.53 87.63 80.9 89.38 87.2 88.81 42.34
Task: TV, Iter:1732/12990; Accuracy on target = 84.01%
T: 96.46 88.2 82.49 63.41 95.31 96.19 85.46 80.65 90.37 89.35 89.59 50.61
Task: TV, Iter:2598/12990; Accuracy on target = 84.38%
T: 96.65 88.6 83.54 62.27 95.25 96.34 87.01 79.97 89.76 91.32 88.62 53.24
Task: TV, Iter:3464/12990; Accuracy on target = 84.73%
T: 96.43 88.49 84.09 63.28 95.57 96.77 86.27 79.38 91.69 91.49 90.63 52.72
Task: TV, Iter:4330/12990; Accuracy on target = 84.87%
T: 96.63 89.09 83.45 57.55 95.61 96.53 86.18 80.65 92.66 91.54 91.64 56.87
Task: TV, Iter:5196/12990; Accuracy on target = 85.13%
T: 96.74 90.39 83.5 59.74 96.25 96.0 85.68 78.92 93.36 93.07 88.81 59.08
Task: TV, Iter:6062/12990; Accuracy on target = 84.71%
T: 97.04 90.59 83.71 57.33 96.03 96.82 81.88 79.82 92.28 92.33 90.49 58.22
Task: TV, Iter:6928/12990; Accuracy on target = 84.81%
T: 96.9 91.97 83.37 55.54 95.84 95.95 82.47 80.35 92.88 92.59 91.29 58.6
Task: TV, Iter:7794/12990; Accuracy on target = 84.18%
T: 96.98 91.42 84.18 48.25 95.65 96.53 82.3 79.72 93.56 93.12 90.53 57.91
Task: TV, Iter:8660/12990; Accuracy on target = 83.79%
T: 96.52 92.75 84.41 44.68 95.74 95.95 81.35 79.55 92.77 93.34 91.67 56.8
Task: TV, Iter:9526/12990; Accuracy on target = 83.87%
T: 97.2 90.5 84.52 41.13 96.35 96.29 86.37 79.88 93.54 91.93 92.33 56.43
Task: TV, Iter:10392/12990; Accuracy on target = 83.60%
T: 97.09 91.83 84.33 39.73 96.1 95.9 85.21 79.32 93.3 93.42 91.9 55.05
Task: TV, Iter:11258/12990; Accuracy on target = 83.73%
T: 97.04 90.96 86.76 40.48 96.5 95.76 87.01 79.18 93.6 93.12 90.93 53.48
Task: TV, Iter:12124/12990; Accuracy on target = 83.71%
T: 97.42 91.91 85.31 40.59 96.76 95.9 85.04 80.12 93.34 92.15 91.45 54.56
Task: TV, Iter:12990/12990; Accuracy on target = 83.59%
T: 96.98 91.31 84.54 39.24 96.4 95.76 85.84 80.05 93.8 93.56 92.23 53.42
The training progress is totally different the one shown in the provided log.
Can you kindly give me some suggestions for reproduction?
Good morning, which hyperparameters should I use to reproduce your Office31 results? I tried with the same of Office-Home but I get results way lower than the ones in the paper.
Hi Shiqi,
While trying to reproduce this work, I found some conflicting package versions. Can you please share the required package versions and dependencies used for implementing this work?
Best,
Anindya
I have another question.
In paper, you have loss for self-regularization and diversity. However, the code only use entropy of prediction as L_div.
Why do you implement like this?
Thanks.
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.