Configuration reference

Every field of net.common.Config with its type and default, generated from the source. Fields can appear at the root of a YAML document, be set with --set field=value or be passed to Config(...). Sectioned keys (data, training, pipeline…) are described in Configuration and The typed feature pipeline.

Sections and cutoff

Field	Type	Default	Note
pipeline	dict | None	None
objective	dict | None	None
optimization	dict | None	None
schema_version	int	1
data	dict	dict
training	dict	dict
parameters	dict	dict
species	dict	dict
layout	dict	dict
layout_rules	dict	dict
codec	dict	dict
reference	dict	dict
cutoff_radius	float	11.0

Representation shape

Field	Type	Default	Note
l_max	int	6
hidden_base_dim	int	64	multiplicity at ℓ = 0
hidden_irreps	str | None	"128x0e+16x0o+8x1e+64x1o+24x2e+8x2o+8x3e+24x3o+" "16x4e+4x4o+4x5e+12x5o+8x6e"
edge_type_emb_dim	int	32	edge type embedding size
emb_use_odd_features	bool	True	use odd parity
edge_encoder_style	str	"rich"	“rich” | “distance”
edge_encoder_use_sh_tensor_square	bool	True
e3layernorm	bool	False

Depth / topology

Field	Type	Default	Note
num_layers_gnn	int	2	unified number of message-passing layers

Model variants

Field	Type	Default	Note
tp_type	str	"separate_weight"
use_self_connection	bool	False
edge_update_node_combine	str	"concat"	“concat” | “sum” | “tensor_product”
edge_update_residual	bool	False	use residual connections in edge update
node_update_message_agg	str	"attention"	“sum” | “average” | “attention”
node_update_attention_scalar_dim	int	64
node_update_attention_heads	int	4
node_update_residual	bool	False	use residual connections in node update
head_use_mlp_log_scale	bool	False	whether to use MLP log scaling in the head
neck_depth	int	2
internal_e3mlp_layers	int	1
head_e3mlp_layers	int	1
head_use_node_embeddings_for_self_edges	bool	True
separate_shifted_self	bool	False
head_use_tensor_square	bool	False
head_pair_mode	str	"split"	“split” | “shared_conditioned”
head_diag_output_scale	float	1.0
head_offdiag_output_scale	float	1.0
head_log_scale_mlp_n_layers	int	1

Non-linearity & norm

Field	Type	Default	Note
e3mlp_variant	str	"film"	“basic” or study-style variants
internal_e3mlp_variant	str | None	"resnormact"
head_e3mlp_variant	str | None	"normact"
e3mlp_output_scale	float	1.0
e3mlp_weight_init_scale	float	1.0
e3mlp_residual_scale	float	0.25
e3mlp_pre_norm	bool	False
e3mlp_norm_eps	float	1e-8
e3mlp_film_hidden_dim	int	64
activation_odd_scalar	str	"tanh"
activation_odd_gate	str	"tanh"
nonlin_kind	str	"normact"	“normact” | “s2act” | “gate_scalars_mlp” | “gate_magnitudes”
activation_scalar	str	"leakyrelu"
activation_gate	str	"softplus"
s2act_res	int	128
norm_kind	str	"component"	for NormActivation: “component” | “norm”

Training

Field	Type	Default	Note
lr	float	0.010803383053097332
max_epochs	int	10000
max_wall_clock_seconds	float | None	12.0 * 3600.0
batch_size	int	1
use_lr_scheduler	bool	True
lr_scheduler_factor	float	0.5
lr_scheduler_patience	int	60
lr_scheduler_min_lr	float	1e-8
lr_scheduler_target	str	"val/hamiltonian_mae"
revert_on_spike	bool	True
revert_monitor	str | None	"val/hamiltonian_mae"
revert_decay_patience	int	4
revert_decay_rate	float	0.5
revert_spike_factor	float	2.0
checkpoint_monitor	str	"val/hamiltonian_mae"

Regularisation

Field	Type	Default	Note
dropout	float	0.0	dropout on all irrep coefficients
l1_reg_coef	float	0.0
l2_reg_coef	float	0.0
init_weights_factor	float	1.0
grad_clip_algorithm	str	"value"
wall_clock_scope	str	"session"
grad_clip_val	float | None	2.0
accumulate_grad_batches	int	1	gradient accumulation steps

Radial basis

Field	Type	Default	Note
n_radial	int	128
radial_layers	Sequence[int]	field( default_factory=lambda: (128, 128, 128) )	e.g. (128,) -> 2-layer MLP
radial_weight_granularity	str	"term"	term (legacy) | copy
pair_conditioned_radial_mlp	bool	False
pair_distance_normalization	str	"off"	off | pair_r0

Additional outputs

Field	Type	Default	Note
enable_forces	bool	False
enable_stress	bool	False
enable_energy	bool	True
enable_num_electrons	bool	True

Training targets

Field	Type	Default	Note
train_target	str	"matrix"	matrix | irreps
partial_train	str | None	None	None | “diag” | “offdiag” (legacy: shifted_self)
train_on_forces	bool	False
train_on_stress	bool	False
train_on_energy	bool	False
train_on_num_electrons	bool	False
matrix_targets	list	["hamiltonian"]
train_observables_on_gt	bool	True
symmetrize_validation_output	bool	True
symmetrize_overlap_targets	bool	True
symmetrize_density_targets	bool	True
symmetrize_output	bool	False	symmetrize matrix outputs
symmetrize_hamiltonian_targets	bool	True
rescale_density_to_num_electrons	bool	False

Loss weighting

Field	Type	Default	Note
loss_l1_fraction	float	1.0	0.0 for L2, 1.0 for L1
matrix_loss_reduction	str	"pair_sum"	pair_sum | pair_mean | elements
loss_coef_observables	float	0.0
allow_zero_observable_loss_control	bool	False
observable_loss_kind	str	"mae"	mse | mae
loss_coef_forces	float	0.0
loss_coef_stress	float	0.0
hamiltonian_envelope_mode	str	"off"	off | normalize_target | multiply_prediction
hamiltonian_envelope_path	str | None	None
hamiltonian_envelope_eps	float	1e-12
loss_weighting_mode	str	"off"
loss_weight_min	float	0.0
loss_weight_max	float	100.0
spectral_loss_enabled	bool	False
train_on_spectral	bool	True
spectral_loss_coef	float	0.0
spectral_loss_kmesh	str	"1x1x1"
spectral_loss_window_ev	float	10.0
spectral_loss_taper_ev	float	2.0
spectral_loss_huber_delta_ev	float	0.1
spectral_loss_kind	str	"huber"	huber | mse | mae
spectral_loss_overlap_psd_cleanup	bool	False
spectral_loss_overlap_jitter	bool	True
spectral_fermi_cache_path	str | None	None
freeze_backbone_train_heads_only	bool	False

Logging

Field	Type	Default	Note
run_name	str	"mandala-run"
experiment_id	str | None	None
ablation_name	str | None	None
ablation_setting	str | None	None
ablation_setting_from	str | None	None
verbosity	int	1
bench_verbosity	int	1
log_partial_gt_observables	bool	True
log_per_irrep_metrics	bool	True
print_per_irrep_metrics	bool	False
log_per_irrep_images	bool	False
log_train_metrics	bool	False
log_per_pair_loss_metrics	bool	True
log_hamiltonian_irrep_contrib_metrics	bool	True
log_hamiltonian_pair_contrib_metrics	bool	True
log_activation_mag	bool	False
log_grad_norm	bool	False
wandb_project	str | None	None
log_every_n_steps	int	1
log_on_step	bool	False	log metrics on step, not just epoch
log_on_epoch	bool	True	log metrics on epoch
log_data	bool	False
log_forward	bool	False
benchmark	bool	False
log_interval	int	10
adaptive_log_interval	bool	True
video_max_atoms	int | None	6

Misc

Field	Type	Default	Note
safety_checks	bool	False	enable strict checks on input data
dtype	torch.dtype	torch.float32	default data type for all layers
device	str	"cuda"	default device for all layers
gpus	int	1	number of GPUs
num_workers	int | None	3
save_dir	str	"checkpoints"	directory to save model checkpoints
log_model	bool	False	whether to log the model to WandB

Caching

Field	Type	Default	Note
snapshot_cache_dir	str | None	None	raw Snapshot .pt cache, if any
dataset_device	str | None	None	keep processed dataset on this device
shuffle_snapshot_load_order	bool	True	randomize cache warmup order
data_split_seed	int	42	fixed seed defining train/validation/test membership
seed	int	42	run seed: initialization, dropout, and training order
epoch_boundary_resume	bool	False	opt-in reproducible segmented runs
precompute_edge_features	bool	True	precompute edge features
radial_embedding_scale	str	"none"
allow_incomplete_dataset	bool	False	tolerate short snapshot lists
apply_cutoff_to_targets	bool	True
graph_source	str	"geometry_cutoff"	geometry_cutoff | target_edges
require_exact_edge_match	bool	True	validate cutoff prefix only; never mutate edges
allow_openmx_positions_box_from_out	bool	False

Hyperopt

Field	Type	Default	Note
tune	str | None	None	hyperparameter tuning (e.g. “ray”, “wandb”)
train_on_irrep_parts	bool	False
compatibility	bool	False

Deeph-e3 specific config

Field	Type	Default	Note
num_block	int	3
r_max	float	8.0
num_basis	int	128
use_sc	bool	True
use_sbf	bool	False