source ~/.bashrc
conda activate mace-scf
which python

mpirun -np 16 python /home/jhm/software/mace-tools/scripts/will_train.py \
        --name="mace_pol_spice_v1" \
        --train_file="../processed_maceoff_scf_spice_1_low_force_filter/train/" \
        --valid_file="../processed_maceoff_scf_spice_1_low_force_filter/val" \
        --test_file "../processed_maceoff_scf_spice_1_low_force_filter/test" \
        --statistics_file="../processed_maceoff_scf_spice_1_low_force_filter/statistics.json" \
        --E0s="{35: -70045.28385080204, 6: -1030.5671648271828, 17: -12522.649269035726, 9: -2715.318528602957, 1: -13.571964772646918, 53: -8102.524593409054, 7: -1486.3750255780376, 8: -2043.933693071156, 15: -9287.407133426237, 16: -10834.4844708122}"\
        --config_type_weights='{"Default":1.0}' \
        --model="Polarizable" \
        --hidden_irreps='64x0e + 64x1o' \
        --loss="energy_forces_atomic_multipoles" \
        --distributed \
        --num_workers=4\
        --num_interactions=2 \
        --r_max=5.0 \
        --energy_weight=40\
        --eval_interval=2 \
        --batch_size=32 \
        --valid_batch_size=32 \
        --max_num_epochs=0\
        --error_table="DensityEnergyRMSE" \
        --swa\
        --start_swa=110\
        --device="cuda"\
        --kspace_cutoff_factor=1.0 \
        --field_dependence_type="biased_local_linear" \
        --energy_key="energy" \
        --forces_key="forces" \
        --atomic_multipoles_weight=10 \
        --swa_atomic_multipoles_weight=1000 \
        --atomic_multipoles_max_l=0 \
        --atomic_multipoles_smearing_width=1.5 \
        --keep_checkpoints \
        --atomic_multipoles_key="mbis_multipoles" \
        --default_dtype="float64" \
        --restart_latest \
        --scf_training_options='{"num_scf_steps":0, "constant_charge":True, "mixing_parameter": 1.0, "use_autograd_forces": True}'