search

google-deepmind / deepmind-research

This repository contains implementations and illustrative code to accompany DeepMind publications
Apache License 2.0
13.08k stars 2.57k forks source link

DGMR Model Code Availability #290

Open jacobbieker opened 2 years ago

jacobbieker commented 2 years ago

Is there any plans to release the model code for the nowcasting DGMR?

shmurai commented 2 years ago

FYI: the pseudocode seems to be released in gs://dm-nowcasting-example-data/pseudocode.zip (in the same bucket as the example datasets).

jacobbieker commented 2 years ago

Awesome! Thanks for finding that. Looks quite complete, only thing I notice is that layers.txt and latent_stack.txt are the same content, so there might be a mistake on copy/paste or something, but that does help a ton!

TQRTQ commented 2 years ago

The code is great, I can understand the details of the paper better, but I found a lot of code missing in layer.txt, such as SNConv3d, ApplyAlongAxis. I have finished a Pytorch version of DGMR code and uploaded it to my GitHub, https://github.com/TQRTQ/DGMR after looking at the pseudo code I found some details that I misunderstood. I want to reproduce this work, can you give me the complete code？please！

bugsuse commented 2 years ago

Awesome! Thanks for finding that. Looks quite complete, only thing I notice is that layers.txt and latent_stack.txt are the same content, so there might be a mistake on copy/paste or something, but that does help a ton!

The pseudocode.zip has been updated, and this issue is fixed.

jacobbieker commented 2 years ago

Thanks for pointing it out! That helps a lot

hyungting commented 2 years ago

Thanks for updating the pseudo code! I am trying to reproduce this work, and I'm wondering if you can generally provide some training advice for me. I followed most of the settings in the paper except several settings that could shorten training period. The training dataset of mine is just simply filtered radar images in dbz (if 40% grids > dbz-> training set). I shortened the predict lead time to 6 frames and set batch size to 45 (due to the limit of my GPU). However, I always got smoothed predictions when the training epoch reached to 300~400, and the loss of discriminators and generators did not decrease. Can you provide some training tips or share your training experience? Thank you!

My code of DGMR in PyTorch version: https://github.com/hyungting/DGMR-pytorch This is my training record of DGMR: https://docs.google.com/presentation/d/13PFavy9GC7Gf6T1k9kFr8G9Xsq_Y2t3WCbgoo0Npc_4/edit?usp=sharing