Configuration

This page summarizes the main configuration objects exposed through the current API.

The canonical configuration definitions live in:

• foreblocks/config.py for ModelConfig and TrainingConfig

ModelConfig

ModelConfig is a lightweight dataclass for architecture-level settings.

Core fields:

Field	Purpose
model_type	selects the backbone family such as LSTM-style vs transformer-style workflows
input_size	feature dimension presented to the model
output_size	output feature dimension
hidden_size	latent width for recurrent or sequence blocks
seq_len	source window length
target_len	forecast horizon
strategy	forecasting strategy such as seq2seq
teacher_forcing_ratio	decoder teacher forcing during training

Extended architecture fields include dim_feedforward, dropout, num_encoder_layers, num_decoder_layers, multi_encoder_decoder, latent_size, and nheads.

It is useful when you want a typed container for model-level settings, even though many examples instantiate modules directly.

TrainingConfig

TrainingConfig controls the training loop.

Core training

Field	Purpose
num_epochs	maximum epoch count
learning_rate	optimizer learning rate
weight_decay	optimizer weight decay
batch_size	default batch size used in examples and helpers
patience	early-stopping patience
min_delta	minimum improvement threshold for early stopping
use_amp	enables automatic mixed precision
gradient_clip_val	gradient clipping threshold
gradient_accumulation_steps	number of mini-batches to accumulate before stepping
l1_regularization	optional L1 penalty
kl_weight	optional KL weight for compatible models

Scheduler and logging

Field	Purpose
scheduler_type	scheduler selector
lr_step_size	step scheduler interval
lr_gamma	step scheduler decay
min_lr	lower learning-rate bound
verbose	trainer logging verbosity
log_interval	batch logging cadence
save_best_model	whether to retain best checkpoint state
save_model_path	optional filesystem path for saved weights
experiment_name	MLTracker experiment name when tracking is enabled

Mixture-of-experts logging

Field	Purpose
moe_logging	enables MoE report collection
moe_log_latency	records latency metrics in MoE logs
moe_condition_name	condition label for segmented MoE analysis
moe_condition_cardinality	category count for condition-aware reports

NAS settings

Field	Purpose
train_nas	enables alternating optimization for architecture parameters
nas_alpha_lr	learning rate for architecture parameters
nas_alpha_weight_decay	weight decay for architecture parameters
nas_warmup_epochs	epochs before alpha updates begin
nas_alternate_steps	alpha-step frequency
nas_use_val_for_alpha	uses validation loss for alpha updates
nas_discretize_at_end	discretizes architecture choices after training
nas_discretize_threshold	threshold used during discretization
nas_log_alphas	logs alpha values during training

Conformal prediction settings

Shared fields:

Field	Purpose
conformal_enabled	master switch
conformal_method	method name such as split, rolling, agaci, enbpi, cptc, or afocp
conformal_quantile	target coverage level
conformal_knn_k	neighborhood size for local conformal methods
conformal_local_window	calibration window for local methods
conformal_aci_gamma	adaptation rate for ACI-style methods
conformal_rolling_alpha	rolling update rate
conformal_agaci_gammas	gamma grid for AgACI
conformal_enbpi_B	bootstrap count for EnbPI
conformal_enbpi_window	rolling window for EnbPI
conformal_tsp_lambda	regularization parameter for TSP
conformal_tsp_window	calibration window for TSP
conformal_cptc_window	state-aware rolling window
conformal_cptc_tau	state filter sharpness
conformal_cptc_hard_state_filter	switches CPTC filtering mode
conformal_afocp_feature_dim	feature dimension for AFOCP
conformal_afocp_attn_hidden	attention hidden size for AFOCP
conformal_afocp_window	context window for AFOCP
conformal_afocp_tau	temperature-like scaling for AFOCP
conformal_afocp_internal_feat_hidden	hidden size of the internal feature network
conformal_afocp_internal_feat_depth	depth of the internal feature network
conformal_afocp_internal_feat_dropout	dropout inside the internal feature network
conformal_afocp_online_lr	online update learning rate
conformal_afocp_online_steps	number of online update steps

Helper methods

TrainingConfig also provides:

• update(**kwargs) for safe field updates
• get_conformal_params() to build the trainer-facing conformal configuration payload

Workflow recipes

Smoke test — no optional features:

config = TrainingConfig(
    num_epochs=5,
    batch_size=16,
    patience=3,
    use_amp=False,
)
trainer = Trainer(model, config=config, auto_track=False)

Preprocessing-heavy run — longer training, AMP, gradient clipping:

config = TrainingConfig(
    num_epochs=50,
    learning_rate=3e-4,
    batch_size=32,
    patience=10,
    use_amp=True,
    gradient_clip_val=1.0,
    scheduler_type="cosine",
    min_lr=1e-6,
    save_best_model=True,
    experiment_name="revin_decomp_baseline",
)

DARTS / NAS run — architecture parameter alternation:

config = TrainingConfig(
    num_epochs=30,
    learning_rate=1e-3,
    batch_size=32,
    train_nas=True,
    nas_alpha_lr=3e-4,
    nas_warmup_epochs=5,
    nas_alternate_steps=1,
    nas_use_val_for_alpha=True,
    nas_discretize_at_end=True,
    nas_log_alphas=True,
)

MoE monitoring run — enable router diagnostics:

config = TrainingConfig(
    num_epochs=40,
    learning_rate=3e-4,
    use_amp=True,
    moe_logging=True,
    moe_log_latency=True,
    moe_condition_name="season",
    moe_condition_cardinality=4,
    experiment_name="moe_router_audit",
    save_best_model=True,
)

Conformal prediction run — split conformal with 90 % coverage:

config = TrainingConfig(
    num_epochs=30,
    batch_size=32,
    conformal_enabled=True,
    conformal_method="split",
    conformal_quantile=0.90,
)

Tracking-oriented production run:

config = TrainingConfig(
    experiment_name="baseline_direct",
    save_best_model=True,
    save_model_path="checkpoints/best.pt",
    use_amp=True,
    gradient_clip_val=1.0,
    scheduler_type="cosine",
)

TimeSeriesHandler settings

Important preprocessor controls include:

• window_size
• horizon
• normalize
• differencing
• detrend
• apply_filter
• apply_ewt
• remove_outliers
• apply_imputation
• generate_time_features
• self_tune

Use the dedicated guide for more detail:

• Preprocessor Guide

Practical guidance

• For first runs, keep configuration small and explicit.
• Disable optional features until the core training loop is working.
• Enable conformal and NAS only after validating your baseline training path.
• Use auto_track=False on Trainer when you do not want MLTracker involved in local smoke tests.

Related pages

• Public API
• Getting Started
• Troubleshooting