search

MeteoSwiss / ldcast

Latent diffusion for generative precipitation nowcasting
Apache License 2.0
111 stars 13 forks source link

DGMR Model Download and Crop Size #3

Closed p3jitnath closed 1 year ago

p3jitnath commented 1 year ago

To test DeepMind's DGMR I did try to run the colab notebook (https://github.com/deepmind/deepmind-research/blob/master/nowcasting/Open_sourced_dataset_and_model_snapshot_for_precipitation_nowcasting.ipynb), and extracting the model from tfhub but it seems it is being served by a GCP Storage Bucket (which I presume will require gsutil but can't seem to get a hold of the endpoint).

For your own evaluations, can you guide on how did you go about downloading the DGMR model and also which crop size did you use (eg. 256x256)?

p3jitnath commented 1 year ago

I think I resolved it by my own.

  1. To download use gsutil cp -r gs://dm-nowcasting-example-data/tfhub_snapshots .
  2. Looking at https://github.com/MeteoSwiss/ldcast/blob/master/scripts/eval_dgmr.py it seems it is 256x256.
jleinonen commented 1 year ago

Yes, the 256x256 model size is used for evaluation.

If you don't want to deal with gsutil it's also possible to download the DGMR model files from the Google Cloud Storage web interface: https://console.cloud.google.com/storage/browser/dm-nowcasting-example-data/tfhub_snapshots. The web interface only allows you to download individual files but since there are only around 10 of them in the model snapshot it's feasible to get the model also that way.