Fix: configs for ENSO - Githubissues

amazon-science / earth-forecasting-transformer

Official implementation of Earthformer

Apache License 2.0

350 stars 60 forks source link

Fix: configs for ENSO #6

Closed gaozhihan closed 1 year ago

gaozhihan commented 1 year ago

Issue #5

Fix the cfg.yaml for training on ENSO.

Change total_batch_size: 32 to total_batch_size: 64 to reproduce the reported results in our paper. total_batch_size: 32 results in suboptimal performance, while total_batch_size: 64 works correctly.
Delete unused config data_seq_len

sxjscience commented 1 year ago

We need to also exchange the metric names in https://github.com/amazon-research/earth-forecasting-transformer/blob/61cab363c8c15d98ecd40643d628303d0234096b/scripts/cuboid_transformer/enso/train_cuboid_enso.py#L533-L534

valid_acc --> valid_corr_nino3.4
valid_weighted_acc --> valid_corr_nino3.4_weighted.

sxjscience commented 1 year ago

Also, change accordingly the test_acc / test_weighted_acc

https://github.com/amazon-research/earth-forecasting-transformer/blob/61cab363c8c15d98ecd40643d628303d0234096b/scripts/cuboid_transformer/enso/train_cuboid_enso.py#L577-L578

gaozhihan commented 1 year ago

We need to also exchange the metric names in

https://github.com/amazon-research/earth-forecasting-transformer/blob/61cab363c8c15d98ecd40643d628303d0234096b/scripts/cuboid_transformer/enso/train_cuboid_enso.py#L533-L534

valid_acc --> valid_corr_nino3.4

valid_weighted_acc --> valid_corr_nino3.4_weighted.

Done

sxjscience commented 1 year ago

Also, we need to add test cases...