Datasets:

deinal
/

njord-sample-data

# Ready made graphs created as follows

python -m neural_lam.create_global_graph \
    --config_path sample_data/global_ocean_1.yaml \
    --name global_cluster_graph \
    --type cluster \
    --plot \
    --levels 2 \
    --grid_to_first_mesh_refinement 5 \
    --connect_disconnected

# ==================================================
# Graph statistics
# ==================================================
#   m2m (level 0): 8456 nodes, 48546 edges
#   m2m (level 1): 2097 nodes, 11640 edges
#   m2m (total): 10553 nodes, 60186 edges
#   mesh_up (level 0->1): 8456 edges
#   mesh_down (level 1->0): 8456 edges
#   g2m: 50804 nodes, 70640 edges
#   m2g: 50804 nodes, 126107 edges
# ==================================================

python -m neural_lam.create_global_graph \
    --config_path sample_data/global_ocean_1.yaml \
    --name global_hierarchical_graph \
    --type hierarchical \
    --plot \
    --levels 2 \
    --connect_disconnected \
    --splits 5

# ==================================================
# Graph statistics
# ==================================================
#   m2m (level 0): 7193 nodes, 40638 edges
#   m2m (level 1): 1812 nodes, 9860 edges
#   m2m (total): 9005 nodes, 50498 edges
#   mesh_up (level 0->1): 7193 edges
#   mesh_down (level 1->0): 7193 edges
#   g2m: 49541 nodes, 67147 edges
#   m2g: 49541 nodes, 125603 edges
# ==================================================

# SeaCast training

python -m neural_lam.train_model \
    --config_path sample_data/global_ocean_1.yaml \
    --num_workers 1 \
    --precision bf16-mixed \
    --model graph_fm \
    --graph global_hierarchical_graph \
    --hidden_dim 16 \
    --hidden_dim_grid 8 \
    --hidden_dim_mesh_nodes 8 \
    --hidden_dim_edge 4 \
    --processor_layers 3 \
    --batch_size 1 \
    --lr 0.001 \
    --ar_steps_eval 2 \
    --val_steps_to_log 1 2 \
    --num_sanity_val_steps 0 \
    --num_nodes 1

# Njord training

python -m neural_lam.train_model \
    --config_path sample_data/global_ocean_1.yaml \
    --num_workers 1 \
    --precision bf16-mixed \
    --model graph_efm \
    --graph global_cluster_graph \
    --hidden_dim 16 \
    --hidden_dim_grid 8 \
    --hidden_dim_mesh_nodes 8 \
    --hidden_dim_edge 4 \
    --processor_layers 3 \
    --batch_size 1 \
    --epochs 10 \
    --lr 0.001 \
    --kl_beta 0 \
    --num_sanity_val_steps 0 \
    --ar_steps_train 1 \
    --ar_steps_eval 2 \
    --val_steps_to_log 1 2 \
    --var_leads_val_plot '{"0":[1,2], "1":[1,2], "2":[1,2], "3":[1,2], "9":[1,2], "15":[1,2]}'

	# Ready made graphs created as follows

	python -m neural_lam.create_global_graph \
	--config_path sample_data/global_ocean_1.yaml \
	--name global_cluster_graph \
	--type cluster \
	--plot \
	--levels 2 \
	--grid_to_first_mesh_refinement 5 \
	--connect_disconnected

	# ==================================================
	# Graph statistics
	# ==================================================
	# m2m (level 0): 8456 nodes, 48546 edges
	# m2m (level 1): 2097 nodes, 11640 edges
	# m2m (total): 10553 nodes, 60186 edges
	# mesh_up (level 0->1): 8456 edges
	# mesh_down (level 1->0): 8456 edges
	# g2m: 50804 nodes, 70640 edges
	# m2g: 50804 nodes, 126107 edges
	# ==================================================

	python -m neural_lam.create_global_graph \
	--config_path sample_data/global_ocean_1.yaml \
	--name global_hierarchical_graph \
	--type hierarchical \
	--plot \
	--levels 2 \
	--connect_disconnected \
	--splits 5

	# ==================================================
	# Graph statistics
	# ==================================================
	# m2m (level 0): 7193 nodes, 40638 edges
	# m2m (level 1): 1812 nodes, 9860 edges
	# m2m (total): 9005 nodes, 50498 edges
	# mesh_up (level 0->1): 7193 edges
	# mesh_down (level 1->0): 7193 edges
	# g2m: 49541 nodes, 67147 edges
	# m2g: 49541 nodes, 125603 edges
	# ==================================================

	# SeaCast training

	python -m neural_lam.train_model \
	--config_path sample_data/global_ocean_1.yaml \
	--num_workers 1 \
	--precision bf16-mixed \
	--model graph_fm \
	--graph global_hierarchical_graph \
	--hidden_dim 16 \
	--hidden_dim_grid 8 \
	--hidden_dim_mesh_nodes 8 \
	--hidden_dim_edge 4 \
	--processor_layers 3 \
	--batch_size 1 \
	--lr 0.001 \
	--ar_steps_eval 2 \
	--val_steps_to_log 1 2 \
	--num_sanity_val_steps 0 \
	--num_nodes 1

	# Njord training

	python -m neural_lam.train_model \
	--config_path sample_data/global_ocean_1.yaml \
	--num_workers 1 \
	--precision bf16-mixed \
	--model graph_efm \
	--graph global_cluster_graph \
	--hidden_dim 16 \
	--hidden_dim_grid 8 \
	--hidden_dim_mesh_nodes 8 \
	--hidden_dim_edge 4 \
	--processor_layers 3 \
	--batch_size 1 \
	--epochs 10 \
	--lr 0.001 \
	--kl_beta 0 \
	--num_sanity_val_steps 0 \
	--ar_steps_train 1 \
	--ar_steps_eval 2 \
	--val_steps_to_log 1 2 \
	--var_leads_val_plot '{"0":[1,2], "1":[1,2], "2":[1,2], "3":[1,2], "9":[1,2], "15":[1,2]}'