from ..config import BeamParam, DeviceConfig, ExperimentConfig, BeamConfig
from ..similarity import SimilarityConfig, TFIDFConfig
[docs]
class TextGroupExpansionConfig(SimilarityConfig, TFIDFConfig):
# "en_core_web_trf"
defaults = {
'chunksize': 1000,
'n_workers': 40,
'mp_method': 'apply_async',
'store_chunk': True,
'store_path': None,
'store_suffix': '.parquet',
'override': False,
'sparse_framework': 'scipy',
}
parameters = [
BeamParam('tokenizer', type=str, default="BAAI/bge-base-en-v1.5", help='Tokenizer model'),
BeamParam('dense-model', type=str, default="BAAI/bge-base-en-v1.5", help='Dense model for text similarity'),
BeamParam('dense_model_device', type=str, default='cuda', help='Device for dense model'),
BeamParam('tokenizer-chunksize', type=int, default=10000, help='Chunksize for tokenizer'),
BeamParam('batch_size', int, 32, 'Batch size for dense model'),
BeamParam('k-sparse', int, 50, 'Number of sparse similarities to include in the dataset'),
BeamParam('k-dense', int, 50, 'Number of dense similarities to include in the dataset'),
BeamParam('threshold', float, 0.5, 'Threshold for prediction model'),
BeamParam('svd-components', int, 64, 'Number of PCA components to use to compress the tfidf vectors'),
BeamParam('pca-components', int, 64, 'Number of PCA components to use to compress the dense vectors'),
BeamParam('pu-n-estimators', int, 20, 'Number of estimators for the PU classifier'),
BeamParam('pu-verbose', int, 10, 'Verbosity level for the PU classifier'),
BeamParam('classifier-type', str, None, 'can be one of [None, catboost, rf]'),
BeamParam('early_stopping_rounds', int, None, 'Early stopping rounds for the classifier'),
]
[docs]
class CatboostConfig(DeviceConfig):
"""
CatBoost configuration with detailed parameter documentation.
References:
- https://catboost.ai/docs/references/training-parameters.html
- https://docs.aws.amazon.com/sagemaker/latest/dg/catboost-hyperparameters.html
"""
defaults = {'objective': None, 'objective_to_report': 'best'}
# CatBoost parameters
parameters = [
BeamParam(
'cb-task',
str,
'classification',
'The task type for the CatBoost model. '
'Default: classification. '
'Options: [classification, regression, ranking]. '
'Determines the type of task and influences the loss function and evaluation metrics.'
),
BeamParam(
'log-frequency',
int,
10,
'The frequency (in iterations) of logging for the CatBoost model. '
'Default: 10. '
'Range: [1, ∞). '
'Affects how often training progress is reported.'
),
# Core parameters
BeamParam(
'loss_function',
str,
'Logloss',
'The loss function for the CatBoost model. '
'Default: Logloss for classification. '
'Options: [Logloss, RMSE, MAE, Quantile, MAPE, Poisson, etc.]. '
'Determines the optimization objective and affects model predictions.'
),
BeamParam(
'eval_metric',
str,
None,
'The evaluation metric for the CatBoost model. '
'Default: Auto-detected based on task type. '
'Options: [Accuracy, AUC, RMSE, MAE, etc.]. '
'Used for evaluating the performance of the model on validation data.'
),
BeamParam(
'custom_metric',
list,
None,
'The custom metric for the CatBoost model. '
'Default: None. '
'Options: [Precision, Recall, F1, etc.]. '
'Provides additional metrics for evaluation.'
),
# Training parameters
BeamParam(
'iterations',
int,
None,
'The number of trees (iterations) in the CatBoost model. '
'Default: 1000. '
'Range: [1, ∞). '
'Higher values may improve performance but can increase training time and risk overfitting.'
),
BeamParam(
'learning_rate',
float,
None,
'The learning rate for the CatBoost model. '
'Default: 0.03. '
'Range: (0.0, 1.0]. '
'Controls the step size at each iteration while moving towards a minimum of the loss function.'
),
BeamParam(
'depth',
int,
None,
'The depth of the trees in the CatBoost model. '
'Default: 6. '
'Range: [1, 16]. '
'Deeper trees can capture more complex patterns but may lead to overfitting.'
),
BeamParam(
'l2_leaf_reg',
float,
None,
'The L2 regularization term on the cost function. '
'Default: 3.0. '
'Range: (0, ∞). '
'Helps prevent overfitting by penalizing large weights.'
),
# Overfitting detection
BeamParam(
'od_pval',
float,
None,
'The threshold for the overfitting detector. '
'Default: None. '
'Range: (0, 1). '
'Stops training if the performance on the validation set does not improve by this value.'
'For best results, it is recommended to set a value in the range [ 1e-10 ; 1e-2 ]'
),
BeamParam(
'od_wait',
int,
None,
'Number of iterations to wait after the overfitting criterion is reached before stopping training. '
'Default: 20. '
'Range: [1, ∞). '
'Prevents premature stopping by allowing continued training for a set number of iterations.'
),
BeamParam(
'od_type',
str,
None,
'The overfitting detection type. '
'Default: IncToDec. '
'Options: [IncToDec, Iter]. '
'Determines how overfitting is detected during training.'
),
# Regularization parameters
BeamParam(
'bagging_temperature',
float,
None,
'Controls the Bayesian bootstrap and helps in reducing overfitting by using random weights. '
'Default: 1.0. '
'Range: [0.0, ∞). '
'Higher values increase randomness, helping to reduce overfitting.'
),
BeamParam(
'random_strength',
float,
None,
'The amount of randomness to use for scoring splits when the tree structure is selected. '
'Use this parameter to avoid overfitting the model.'
'Default: 1.0. '
'Range: [0.0, ∞). '
'Adds randomness to scoring splits, helping prevent overfitting.'
),
# Feature processing
BeamParam(
'border_count',
int,
None,
'The number of splits for numerical features (max_bin). '
'Default: The default value depends on the processing unit type and other parameters: '
'CPU: 254 '
'GPU in PairLogitPairwise and YetiRankPairwise modes: 32 '
'GPU in all other modes: 128 '
'Range: [1, 65535]. '
'Affects the granularity of feature discretization; higher values can improve accuracy but increase complexity.'
),
BeamParam(
'feature_border_type',
str,
None,
'The feature border type. '
'Default: GreedyLogSum. '
'Options: [Median, Uniform, UniformAndQuantiles, MaxLogSum, GreedyLogSum, MinEntropy]. '
'Determines how feature borders are selected, impacting split decisions.'
),
BeamParam(
'per_float_feature_quantization',
str,
None,
'The per float feature quantization. '
'Default: None. '
'See: https://catboost.ai/en/docs/references/training-parameters/quantization. '
'Allows custom quantization for specific features.'
),
# Advanced tree options
BeamParam(
'grow_policy',
str,
None,
'Defines how to perform greedy tree construction. '
'Default: SymmetricTree. '
'Options: [SymmetricTree, Depthwise, Lossguide]. '
'Determines the strategy for tree growth, affecting complexity and interpretability.'
),
BeamParam(
'max_leaves',
int,
None,
'The maximum number of leaves in the resulting tree. '
'Default: None. '
'Range: [2, 64]. '
'Applicable for Lossguide grow policy; limits the complexity of the tree.'
),
# Sampling and randomness
BeamParam(
'rsm',
float,
None,
'Random subspace method for feature selection. '
'Default: 1.0. '
'Range: (0.0, 1.0]. '
'Percentage of features used at each split selection, allowing randomness in feature selection.'
),
BeamParam(
'sampling_frequency',
str,
None,
'Frequency to sample weights and objects during tree building. '
'Default: PerTree. '
'Options: [PerTree, PerTreeLevel]. '
'Determines how often samples are drawn during training.'
),
BeamParam(
'bootstrap_type',
str,
None,
'The bootstrap type. '
'Default: Bayesian. '
'Options: [Bayesian, Bernoulli, No, MVS]. '
'Controls how samples are drawn for training, affecting robustness and variance.'
),
# Leaf estimation
BeamParam(
'leaf_estimation_iterations',
int,
None,
'The number of iterations to calculate values in leaves. '
'Default: 1. '
'Range: [1, ∞). '
'Higher values improve accuracy at the cost of increased training time.'
),
BeamParam(
'leaf_estimation_method',
str,
None,
'The method used to calculate values in leaves. '
'Default: Newton. '
'Options: [Newton, Gradient]. '
'Determines the approach for estimating leaf values, affecting convergence speed and accuracy.'
),
# Logging and output
BeamParam(
'snapshot_interval',
int,
600,
'The snapshot interval for model saving [in seconds]. '
'Default: 600. '
'Range: [1, ∞). '
'Controls how often model snapshots are saved, useful for resuming training.'
),
BeamParam(
'boosting_type',
str,
'Plain',
'Controls the boosting scheme. '
'Default: Plain. '
'Options: [Ordered, Plain]. '
'Ordered is used to eliminate the effect of a training set order.'
),
BeamParam(
'allow_const_label',
bool,
False,
'Allows training on a dataset with constant labels. '
'Default: False. '
'Useful for experimentation or testing.'
),
# Training parameters
BeamParam(
'auto_class_weights',
str,
None,
'Automatically calculates class weights for imbalanced datasets. '
'Default: None. '
'Options: [Balanced, SqrtBalanced, None].'
),
# Regularization parameters
BeamParam(
'l1_leaf_reg',
float,
None,
'L1 regularization term on weights. '
'Default: 0.0. '
'Range: (0, ∞). '
'Helps prevent overfitting by penalizing large weights.'
),
# Feature processing
BeamParam(
'one_hot_max_size',
int,
None,
'Maximum size of the categorical feature for one-hot encoding. '
'Default: 2. '
'Range: [1, ∞). '
'Larger sizes use a more efficient embedding.'
),
# Advanced tree options
BeamParam(
'min_data_in_leaf',
int,
None,
'Minimum number of samples per leaf. '
'Default: 1. '
'Range: [1, ∞). '
'Controls leaf size and can affect overfitting and generalization.'
),
# Sampling and randomness
BeamParam(
'bagging_fraction',
float,
None,
'Fraction of samples to use in each iteration. '
'Default: 1.0. '
'Range: (0.0, 1.0]. '
'Controls randomness and variance.'
),
# Leaf estimation
BeamParam(
'leaf_estimation_backtracking',
str,
None,
'Backtracking type used for leaf estimation. '
'Default: AnyImprovement. '
'Options: [No, AnyImprovement, Armijo]. '
'Affects convergence and accuracy.'
),
]
[docs]
class CatboostExperimentConfig(CatboostConfig, ExperimentConfig):
defaults = {'project': 'cb_beam', 'algorithm': 'CBAlgorithm'}
