google / mentornet Goto Github PK

Hi @roadjiang , Thank you for contributing this amazing codebase. I am wondering if you can share the codes to generate the noisy data? I would like to try different noise ratio on CIFAR10 and CIFAR100 and will really appreciate your help1

您好，麻烦问下您代码中哪个地方是您论文中所述的“ In Step 6, the MentorNet parameter Θ is updated to adapt to the most recent
model parameters of StudentNet. In experiments, we update
Θ twice after the learning rate is changed. Each time, a datadriven curriculum is learned from the data generated by the
most recent w using the method discussed in Section 3.1”？
谢谢。

Can you tell me where the clean label come from in the .csv fiel? Is it the labels of small clean dataset? The noisy labels is the prediction of pre-trained model on noisy dataset.

I am curious about how the # of parameters are computed (https://arxiv.org/pdf/1712.05055.pdf).

Seems like the code here uses some sort of WideResNet-101-10, but since WideResNet-28-10 has around 36.5M params, a network that is 101 layers would give almost 1.5B parameters right?

This is almost 2x larger than your reported numbers. Or did you use a smaller architecture in the paper?

In the supplementary of paper, it writes that as CIFAR-100 and CIFAR-10 have the different number of classes, to apply a MentorNet, we fix the class label to 0. It's not clear which label is fix to 0, because there are two labes for samples, i.e., clean labels and noisy labels.

epoch_embedding = tf.get_variable( 'epoch_embedding', [100, epoch_embedding_size], trainable=False)
Because epoch_embedding is not tainable, how to initialize this variable when train mentornet from zero?

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

1. Does it also work on tabular data?
2. Is the code tailored to the datasets used in the paper or can one apply it to any data?
3. Is it possible to identify the noisy instances (return the noisy IDs or the clean set)?

Thanks!

I want to know how the csv file used to train mentor-dd is generated.
My understanding is to train the baseline model with a clean tag dataset and use the Corrupted Labels dataset to calculate the loss to get the csv file. Can you tell me the details of generating a csv file?

This is a question regarding the paper, not the code.
I tried to parse through your code, but as I am not familiar with TF, there are more confusions.

As far as I understand, MentorNet (DD) is only trained with a small clean set, right?
If the label spaces are matched between clean set and large noisy set, it is pretty clear.

My question is, as there are different number of classes between CIFAR10 and CIFAR100, how can you train label embedding layer in CIFAR10 and deploy in CIFAR100?

Thank you :D