ecmwfcode4earth / deepr Goto Github PK

Add some extra validation plots:

• Sampler. Implement validation sampler to reduce the validation time. Needed for DP due to its high inference cost for generating one sample.

• Time. New maps:

  • Error maps and correlation maps with respect to the hour of the day and the month of the year.

• Sea vs Land. Load land mask in test set. New plots:

  • Specific time series for land & sea.
  • Specific histograms for land & sea.

• Orography. Load orography & land mask in the test set. New plots:

  • Compute error plots by orography (only in land). As it is a continuous variable, it may be useful to use bins for categorizing it.

Add two specific figures for DP validation:

• Generation GIFF: Figure with 2 plots. Left plot: a ERA5 sample (static). Right plot: the prediction X_t from t=0,...,1000 to visualize the generation process starting from noise. Another option is to consider only the right plot for the GIF and print the correspoding ERA5 sample in a separated PNG
• Move samples comparison: Move the recurrent save_samples func from diffusion_trainer.py to the evaluation of the model (due to its computation cost).