GAN-RNN_Timeseries-imputation
GAN-RNN_Timeseries-imputation copied to clipboard
About noise input for GAN
Hi Ivan, excellent work and thanks for making the codes open! I've read your codes and the markdown for how it works. It seems that there is no noise input for GAN and the partial GAN according to your implementation. I'm not an expert in GAN and time series imputation task, but is it common to do so?
Hi, thank you for your interest in my project! Yeah you're right, it's not a complete, canonical GAN. In this case I didn't add any random noise even though that is what you normally have in GANs. The reason is that while classical GANs generate observations from pure noise, in my case the generation is produced from an already existing (deteriorated) trend. So it's not a 100% pure generative task.
I wrapped my head around this problem, I'm thinking about ways to add randomness in one way or another . Also, I don't know if adding it could improve the model's performance.
Food for thought! Do you have an opinion about this?
Hi Ivan, thanks so much for your clarification. By "deteriorated trend", you mean the time series masked by artificial missing, right? Since I actually began implementing GAN for a similar task, I think it deserves further experiments to discuss the performance difference. But thanks again for your easy-to-follow codes and illustrations. It makes a good start for me and others!
Sorry for the delay!
Yes, by "deteriorated trend", I mean the time series masked by artificial missing. Please share your work as well, I'm very interested. Let me (us) know if you make something cool!
Hi everyone,
I'm really interested in this GAN RNN, I am reading code and trying to figure out each step/method's functions, but not fully understand and nowhere I can find documents. Therefore, I'm writing here in the hope that you can give me your contact so that I can discuss hopefully
At present, I don't know this block
Not sure why we have to prepare that kind of numpy data. What can it reveal?
Thanks
Hi guys, 😃 SAITS may be useful to you on the time-series imputation problem. You can try SAITS with PyPOTS https://github.com/WenjieDu/PyPOTS with only few lines of code. Please take a look if you're interested.
And thank you @IvanBongiorni very much for creating this repo that gives us a space to discuss here 👍