xjtushujun/meta-weight-net

NeurIPS'19: Meta-Weight-Net: Learning an Explicit Mapping For Sample Weighting (Pytorch implementation for noisy labels).

meta-weight-net's Issues

有关GPU的问题

老师你好，请问这个Meta-resnet和Meta-vnet可以用多块GPU吗

对于训练过程里的代码有个疑问150~161

optimizer_c中优化的是vnet的参数，但是损失l_g_meta中我没有看到vnet网络的参与，我有点想不清楚网络是如何迭代的

how to draw the accuracy curve

Hi，I have a question When I repeated the experiment in your paper. I was not clear how to draw the accuracy curve of the training set.
Thanks.

About the effectiveness

Thank you for your excellent work!

Here, we raised some questions about the effectiveness (time & memory) of Meta-Weight-Net：

Compared to Learn-to-Reweight (Ren, 2017), how about the cost of running time and GPU memory per training step?

Can the training process's time efficiency be improved by updating Meta-Weight-Net every several steps (rather than updating every step)? Will this affect model's performance?

Is it possible to achieve multi-GPU parallelism (based on Pytorch)?

Thanks very much~

About the details of learning rate

There is a sentence in the appendix: "With batch normalization, we effectively cancel the learning rate of Meta-Weight-Net, and it works well with a fixed learning rate. "

I'm not sure what it is about. Would you please give an explanation in detail? Does it mean we don't need to fine-tune the learning rate of meta networks because of BN?

请问一下能否用Adam作为optimizer_model的优化算法？

请问一下能否用Adam作为optimizer_model的优化算法？

tabular data/ noisy instances

Hi,
thanks for sharing your implementation. I have two questions about it:

1. Does it also work on tabular data?
2. Is it possible to identify the noisy instances (return the noisy IDs or the clean set)?

Thanks!

关于模型的问题

首先，非常感谢你们开源这么优秀的工作！

我有一个疑惑，为什么作者重写了整个模型。

基于MetaModule重写了所有卷积，线性和批量归一化的目的是什么呢？为什么不直接使用torch.nn模块里的模型（如，nn.Conv2d），就如同pytorch官方实现的resnet等。

计算图问题

Can you provide the code for imbalanced dataloader?

Hi, I found there is no implement for imbalanced dataset. Could you please provide it or give a reference link?

数据集划分

请问实验中用到的cifar-10和-100，是怎么划分训练集或者验证集的呢？文中的meta-learning是用的episodic方式将数据组织成n-way-k-shot的吗？文中的实验结果是多少n，多少k呢？能告知一下吗？

The accuracy of BaseModel is 88.5 when the noise rate is 0.4?

I trained a WRN-28-10 network on cifar-10 with noise rate of 0.4 under uniform noise following the setting in the paper for a total of 40 epochs, but the accuracy of BaseModel is 88.5, which is really high compared with the results in Table 2. I don't know what the problem is.

请问本文的方法和L2RW的比较

文章提到L2RW "might lead to unstable weighting behavior during training and unavailability for generalization"
请问这是为什么？依据是什么？

关于baseline精度问题

请问baseline 的参数设置和其他方法的设置一样吗，为什么baseline在噪声40%的结果偏高呢，大约0.8左右

想问问大佬那个Figure 1(f)的权重跟Loss的分布图是怎么画的。

我在文本层次多标签分类任务上复现了文章中的meta-weight-net，但是感觉效果并没有提升，对比我的baseline只上升了0.2%，我想画出权重跟loss的分布图分析下原因。