Tsing

https://www.linkedin.com/in/tsingz/ [email protected]

Shanghai Jiao Tong University (SJTU) Shanghai, China

Results 27 comments of


                                            Tsing

In the pathological Non-IID setting, the samples distribution on clients may be unbalanced even the `balance` is True.

Hi, @bird-two. Thank you for your valuable comments! The problem you mentioned does exist when generating the pathological Non-IID setting with the unbalanced raw dataset. As the statistical heterogeneity in...

from opacus.dp_model_inspector import DPModelInspector

Hi, @qq648545022 I mentioned "By using the package [opacus v0.15](https://github.com/pytorch/opacus/releases/tag/v0.15.0)" in the README.md, please check it.

Hi, I can't see local representations learning in main_lg.py. It just records the best local acc and the best local model, then updates the global model.

> I think this code can't reduce the number of parameters. The code does not implement the ideas in your paper. Is there a error with my understanding? @wardseptember After...

Questions about the test_metrics() functions and the partition between train and test.

For question one (evaluate function): This is a platform for research, not for production currently. You can modify any of the code as you want. If testing on all the...

A question about backpropagation

I also meet the same question as @cuicathy.

A question about backpropagation

Instead of using `OLE.py`, I find that PyTorch can compute the nuclear norm by `loss = torch.norm(X, p='nuc')`.

Share some fine-tuned hyperparameter settings

是的，`-gr`的设置没有严格要求，如果算法能在1000轮内收敛，那么`-gr 1000`和`-gr 2000`最终结果是一样的

resnet18训练Cifar100数据集准确率很低

这里的”对于Cifar100数据集即使将客户端数量修改为1“指的是用dataset文件夹下的代码切分数据集的时候，就设置为1个客户机吗？我们这边默认设置数据量：tran set: test set=3:1，而且是先将train set和test set混合后再切分为3:1的。原始数据集tran set: test set=5:1，相比而言，训练集数据量更多，更不容易过拟合。对于模型训练而言，训练数据量至关重要，如果希望跟集中式训练对比，那么请修改相应代码，保证这方面的一致性。

resnet18训练Cifar100数据集准确率很低

很抱歉，超参数选择上，我这边给不了太多的建议。因为我在做研究的过程中，本着“尽可能不调参”、“尽可能在各种任务上只使用一套超参数”的理念，去设计联邦学习算法。

(New data partition strategy) Extended Dirichlet strategy by combining Pathological heterogeneous setting and Practical heterogeneous setting in pFL.

You can contribute to our project by submitting a pull request that adds the Extended Dirichlet strategy. We may add it when we have free time.

1
2
3
›