Trainer Node

The Trainer node is the core of model training in MLOps Desktop. It supports three modes: training new models, loading pre-trained models, and hyperparameter tuning with Optuna.

Overview

Property	Value
Type	Processing node
Inputs	DataFrame (from DataLoader or DataSplit)
Outputs	Trained model
Library	scikit-learn
Modes	Train, Load, Tune

Operating Modes

The Trainer node has three modes, selected via toggle buttons:

Train

Train a new model from scratch.

Configuration:

• Model Type — Select from 12 algorithms
• Target Column — Column to predict
• Test Split — Ratio for train/test split (if not using DataSplit node)

Best for initial model development and quick experiments.

Load

Load a pre-trained model from disk.

Configuration:

• Model File Path — Path to .joblib, .pkl, or .pickle file

Caution

A security warning appears before loading pickle files, as they can execute arbitrary code.

Best for using models trained elsewhere or resuming work.

Tune

Automatically find optimal hyperparameters with Optuna.

Configuration:

• Model Type — Model to tune (Linear Regression disabled—no tunable params)
• Target Column — Column to predict
• Tuning Config — Search strategy, trials, CV folds, search space

Best for maximizing model performance before production.

Supported Models

Classification Models (6)

Model	Description	Key Hyperparameters
Logistic Regression	Linear classifier, interpretable	C, max_iter
Random Forest Classifier	Ensemble of decision trees	n_estimators, max_depth, min_samples_split
Gradient Boosting Classifier	Sequential boosting, high accuracy	n_estimators, learning_rate, max_depth
SVM (SVC)	Support vector machine	C, kernel, gamma
KNN Classifier	Distance-based classification	n_neighbors, weights, metric
MLP Classifier	Neural network	hidden_layer_sizes, alpha, learning_rate_init

Regression Models (6)

Model	Description	Key Hyperparameters
Linear Regression	Simple linear model	None (no tuning)
Random Forest Regressor	Ensemble for regression	n_estimators, max_depth, min_samples_split
Gradient Boosting Regressor	Boosted trees for regression	n_estimators, learning_rate, max_depth
SVM (SVR)	Support vector regression	C, kernel, gamma
KNN Regressor	Distance-based regression	n_neighbors, weights, metric
MLP Regressor	Neural network for regression	hidden_layer_sizes, alpha, learning_rate_init

Hyperparameter Tuning

When Tune mode is selected, click the tuning config button to open the TuningPanel.

Search Strategies

Strategy	Description	Best For
Bayesian (TPE)	Tree-structured Parzen Estimator, learns from past trials	Most cases (default)
Random	Uniform random sampling	Baseline comparison
Grid	Exhaustive enumeration of all combinations	Small, discrete spaces

Tuning Configuration

Setting	Range	Default	Description
Number of Trials	1-1000	50	How many configurations to try
CV Folds	2-10	3	Cross-validation folds
Scoring Metric	varies	accuracy/r2	Metric to optimize

Search Spaces by Model

Each model has predefined search ranges:

Random Forest:

n_estimators: 50-300 (step 50)
max_depth: [null, 10, 15, 20, 30]
min_samples_split: 2-10 (step 2)
min_samples_leaf: 1-4 (step 1)

Gradient Boosting:

n_estimators: 50-300 (step 50)
learning_rate: 0.01-0.3 (log scale)
max_depth: 3-8 (step 1)
subsample: 0.7-1.0 (uniform)

SVM (SVC/SVR):

C: 0.1-100 (log scale)
kernel: [rbf, linear, poly]
gamma: [scale, auto]

KNN:

n_neighbors: 3-21 (step 2)
weights: [uniform, distance]
metric: [euclidean, manhattan, minkowski]

MLP Neural Network:

hidden_layer_sizes: [(50,), (100,), (100,50), (100,100)]
alpha: 0.0001-0.1 (log scale)
learning_rate_init: 0.0001-0.1 (log scale)
max_iter: 200-1000 (step 100)

Scoring Metrics

Classification:

• Accuracy
• F1 Score
• Precision
• Recall
• ROC AUC

Regression:

• R² Score
• Neg MSE
• Neg MAE
• Neg RMSE

Automatic Preprocessing

The Trainer automatically handles:

• Missing Values — Numeric columns filled with median, categorical with mode
• Categorical Encoding — Label encoding for all categorical columns
• ID Column Filtering — Drops columns like id, index, name, ticket, cabin
• High-Cardinality Columns — Drops columns with >50 unique values

Connections

Direction	Node Types
Input from	DataLoader, DataSplit
Output to	Evaluator, ModelExporter

Typical pipeline:

DataLoader → DataSplit → Trainer → Evaluator

Trials Panel

When tuning, results appear in the Trials tab:

Column	Description
Trial #	Trial number
Score	Cross-validation score
Parameters	Hyperparameter values used
Duration	Time taken
Status	Complete, Pruned, or Failed

The best trial is highlighted with a star icon.

Common Issues

”Linear Regression cannot be tuned”

Linear Regression has no tunable hyperparameters. Use Train mode instead, or select a different model.

”Optuna not installed”

Install Optuna:

pip install optuna

Tuning is slow

• Reduce number of trials
• Use Random search instead of Grid
• Reduce CV folds (minimum 2)
• Use a faster model (Logistic Regression vs MLP)

“Target column not found”

Check that the column name matches exactly (case-sensitive). Use the DataLoader preview to verify column names.

Related Nodes

• DataLoader — Load training data
• DataSplit — Split into train/test sets
• Evaluator — Evaluate trained models
• ModelExporter — Export models