search

tim-learn / SHOT

code released for our ICML 2020 paper "Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation"
MIT License
437 stars 78 forks source link

use domainnet dataset #28

Closed Chy-dev249 closed 2 years ago

Chy-dev249 commented 2 years ago

Hi ,thanks for your excellent work. Have you ever try to work on the domainNet dataset ? SInce I can't get satisfied result by changing the dataset only.

tim-learn commented 2 years ago

Hi ,thanks for your excellent work. Have you ever try to work on the domainNet dataset ? SInce I can't get satisfied result by changing the dataset only.

Hi, maybe you need to change the learning rate and adjust the weights of different loss objectives by yourself since DomainNet owns an extremely imbalanced label distribution.

Best

Chy-dev249 commented 2 years ago

Hi, thanks for your response. But I still want to know whether the kmeans is still work in domainnet. (From my experiment result, I found it's not work well ) Since the number of classes is 345, the variable of k in kmeans has better to be a smaller one, like 8. Is it necessary to change another method to optimize the pseudo lable during the training process? Do you have any suggestions? Thank you very much！ Best

tim-learn commented 2 years ago

Hi, thanks for your response. But I still want to know whether the kmeans is still work in domainnet. (From my experiment result, I found it's not work well ) Since the number of classes is 345, the variable of k in kmeans has better to be a smaller one, like 8. Is it necessary to change another method to optimize the pseudo lable during the training process? Do you have any suggestions? Thank you very much！ Best

You can try the updated code for DomainNet again.

Externalhappy commented 1 year ago

I have the same problem too. Even when changing the learning rate or adjusting the weight of the loss function, the algorithm may not work or give satisfactory results on large datasets. Have you succeeded solving this issue, @Chy-dev249 ?