search

google-deepmind / graphcast

Apache License 2.0
4.5k stars 562 forks source link

Help regarding prediction of target variables for the year 2022 #94

Closed ElNino9495 closed 1 week ago

ElNino9495 commented 1 week ago

I have been trying to run an experiment where I'm trying to predict the target variables of graphcast for the year 2022. I am fetching my input data from CDS. The time taken to calculate forcings for a single day is taking me a very long time(2 hours plus). I don't think I'm following the right approach. How can i run this experiment quickly? For compute, I have a server which has access to 5 NVIDIA 100 GPUs. Any help would be greatly appreciated!

ElNino9495 commented 1 week ago

I keep getting this error while trying to make predictions using the /data4/home/rohitsuresh/graphcast_/model/params/params-GraphCast_operational-ERA5-HRES_1979-2021-resolution_0.25-pressure_levels_13-mesh_2to6-precipitation_output_only.npz model. My inputs, targets and forcings are all in the same resolution as well

The error: ValueError: 'grid2mesh_gnn/ ~_networks_builder/encoder_nodes_grid_nodes_mlp/~/linear_0/w' with retrieved shape (184, 512) does not match shape=[186, 512] dtype=dtype(bfloat16)

ElNino9495 commented 1 week ago

I was adding precipitation as an input.