I'm glad to see this reworking.

[TQRTQ]

I am a graduate student majoring in deep learning. I am also reproducing this Nature paper and have finished the code work. I want to talk to you about the code. This is my code address：https://github.com/TQRTQ/DGMR

[jacobbieker]

Nice! Your implementation looks quite good, does it train correctly? This repo's implementation still gives NaN values in the Sampler.

[TQRTQ]

I have not made use of the data provided in the paper to conduct experiments, and I have encountered some difficulties in processing the data in the paper. So I randomly initialized some data, and the code worked fine.
But tention is my spatial intention, you is self-attention. I'll read your code carefully and learn about it. In addition, I would like to ask you about data preprocessing

[jacobbieker]

Yeah, they released the pseudocode, which is now also in this repo. So my attention is copied from theirs. Currently, I haven't worked with their data yet, so its just been random data in the unit tests. I'll probably start with their data soonish though to make sure it works, and with porting over their weights.

[hyungting]

Hi! I'm a master student, my thesis is about radar precipitation nowcasting with machine learning methods. I'm wondering if you are still working on training DGMR. I'm now trying to train DGMR but the predictions are smoothed and the discriminator losses did not decrease that much. If you are still working on this project, maybe we can exchange some training experiences. I've also implemented DGMR: https://github.com/hyungting/DGMR-pytorch

[jacobbieker]

I am still working on it, the implementation should be fairly complete, it is nearly a 1 to 1 copy of their pseudocode, including for training. I haven't trained the model yet on the dataset they released, but will try to soon.

[peterdudfield]

@all-contributors please add @TQRTQ for question

[allcontributors]

@peterdudfield

I've put up a pull request to add @TQRTQ! 🎉