Training scripts for CPI dataset

[nnsriram97]

Hey, is training and testing scripts used for the CPI dataset available? Thanks

[os1a]

Hi, Thanks for your interest in our work.

We are planning to release an updated version (using pytorch) to train the framework which will make it easier for people to train and test the framework.

Meanwhile, feel free to raise more questions if you want to implement it yourself.

I will keep this issue open until we have the updated version.

Best,

[nnsriram97]

Thanks, looking forward to your PyTorch implementation.

May I know how did you avoid mode collapse on the fitting network? All the soft assignments assigning predicted hypotheses to one mode of the GMM. Found this while implementing, any insights on training the fitting network to avoid this issue?

[os1a]

Hi @nnsriram97,

The main idea from the paper is to overcome the mode collapse which is usually the case when you train end-to-end. In our work, we first train the sampling network to generate diverse hypotheses and then the fitting on top of the sampling. Note that we fix the sampling network when training the fitting network. In other words, the fitting network has to find the parameters of GMM with different modes to fit the diverse set of hypotheses. Only at the end, we do joint training of the two networks.

So to summarize, we have three stages:

• train the sampling network with EWTA.
• train the fitting network with NLL while keeping the weights of the sampling network fixed.
• train jointly both of them with NLL.

[seoho-kang]

Thanks for the great work! How long did you take to train each stage(time or steps?) Also, can you let us know when Torch implementation(Training, Testing) would be available? Really looking forward to Torch implementation :)

[os1a]

@droneRL2020

You can check the Figure 3 in our supplementary material of the arXiv paper: https://arxiv.org/pdf/1906.03631.pdf

All stages of the EWTA are trained equally, this means that if we train for 150k iterations using EWTA and we have 5 stages, then every stage is trained for 30k.

I cannot give exact dates for the torch implementation but we hope to do it within 5-7 weeks.

Best,

[seoho-kang]

I'm currently trying to train the model. If I have more questions I will leave it here. Thanks for the great work!

[rafalk342]

@os1a How is the progress going with the pytorch implementation?

[os1a]

Hi @rafalk342

Thanks for your interest in our code,

Unfortunately we are still waiting for an approval from our business partner, but we are working on it and hopefully we can publish it before the end of the year.

I will post here when we have it ready.

[Shaluols]

Is there any progress on the training script release? Or has somebody replicated the training process successfully?

[os1a]

Hi @Shaluols,

Thanks for your interest in our codebase. Unfortunately publishing the source code for training will take longer than expected. The main problem is the transfer between frameworks. The original code we had was trained on Tensorflow with two-privately libraries used for data reading and augmentation.

I will update you as soon as we have it ready. Sorry for the delay,

Best,

[Shaluols]

@os1a Thanks for the update~

[mprzewie]

Hi all!

I currently work on a research project related to this work. We have adapted the parts of code from this repository to generate the CPI dataset files in the identical format to the SDD data so that one can plug-and-play with it.

https://github.com/maciejzieba/regressionFlow/tree/master/cpi_generation

[os1a]

Thanks @mprzewie for sharing the code,

Your work is pretty good and I liked it.

Best,

[ThierryDeruyttere]

@os1a Any update on releasing the training code? Thanks!