5.1. Classification Dataset Format

This guide explains how to format datasets for time series classification tasks.

5.1.1. Directory Structure

Classification datasets use a classes/ folder where each subfolder represents a class:

my_dataset/
└── classes/
    ├── class_A/
    │   ├── sample1.csv
    │   ├── sample2.csv
    │   └── sample3.csv
    ├── class_B/
    │   ├── sample1.csv
    │   └── sample2.csv
    └── class_C/
        └── sample1.csv

Key points:

Folder names become class labels
Each CSV file is one sample (or multiple samples if using windowing)
All files should have the same number of columns

5.1.2. Data File Format

Headerless Format (Simple)

Just numeric values, one measurement per row:

Headered Format (Recommended for Multi-Variable)

First row contains column names:

channel_x,channel_y,channel_z
0.523,0.112,-0.234
0.612,0.098,-0.198
0.498,0.145,-0.267
...

Time Column Handling

Any column containing “time” (case-insensitive) is automatically dropped:

Time,value1,value2
0.001,0.523,0.112
0.002,0.612,0.098
...

The “Time” column will be removed, leaving only value1 and value2.

5.1.3. Supported File Types

Early stopping is on by default in 1.5.0 for all tasks, so a clean validation split matters: see Early Stopping.

Extension	Description
`.csv`	Comma-separated values (most common)
`.txt`	Tab or space-separated text
`.npy`	NumPy array (binary, faster loading)
`.pkl`	Pickled pandas DataFrame

5.1.4. Annotations (Optional)

You can optionally provide train/val/test splits using annotation files:

my_dataset/
├── classes/
│   └── ...
└── annotations/
    ├── file_list.txt              # All files (auto-generated if missing)
    ├── instances_train_list.txt   # Training files
    ├── instances_val_list.txt     # Validation files
    └── instances_test_list.txt    # Test files (optional)

File List Format

Each annotation file lists relative paths, one per line:

# instances_train_list.txt
class_A/sample1.csv
class_A/sample2.csv
class_B/sample1.csv

If the annotation files are missing, ModelMaker auto-generates them under annotations/ using split_type and split_factor from the config (see Data Splitting).

ModelMaker looks for each split list in this order: <split>_list.txt in the dataset root, annotations/<split>_list.txt, then annotations/<annotation_prefix>_<split>_list.txt (default prefix instances). The split names are train, val and test. If any one split list is missing, all splits are regenerated.

5.1.5. Configuration

dataset:
  enable: True
  dataset_name: 'my_classification_data'
  input_data_path: '/path/to/my_dataset'  # or URL to .zip
  data_dir: 'classes'              # Default
  annotation_dir: 'annotations'    # Default (optional)
  split_type: 'amongst_files'
  split_factor: [0.6, 0.3, 0.1]    # train, val, test

data_processing_feature_extraction:
  variables: 3                     # Number of data columns

Note

Zip file structure requirement: When using a zip file as input_data_path, the zip must contain the classes/ directory immediately inside it (at the top level). Do not add an extra directory level such as dataset_name/classes/ inside the zip. The same applies to the optional annotations/ directory.

5.1.6. Dataset Splitting Modes

The split_type parameter controls how ModelMaker divides data into train, validation, and test sets when the annotations/ folder is not provided.

amongst_files (default)

Entire files are assigned to train, validation, or test sets. For example, with 10 files (each having 100 rows) and the default split_factor: [0.6, 0.3, 0.1]:

6 files go to training (each retains all 100 rows)
3 files go to validation
1 file goes to test

Use this mode when each file represents a distinct experiment or recording session.

within_files

Each file is split internally into train, validation, and test portions. For example, with 10 files (each having 100 rows):

All 10 files appear in every split
Training portion: first 60 rows of each file
Validation portion: next 30 rows of each file
Test portion: last 10 rows of each file

Use this mode when files contain long continuous sequences that can be safely split at arbitrary points.

dataset:
  split_type: 'within_files'       # or 'amongst_files' (default)
  split_factor: [0.6, 0.3, 0.1]    # train, val, test proportions

5.1.7. Example: 3-Class Vibration Data

Dataset structure:

vibration_dataset/
└── classes/
    ├── normal/
    │   ├── run1.csv
    │   ├── run2.csv
    │   └── run3.csv
    ├── fault_type_A/
    │   ├── fault1.csv
    │   └── fault2.csv
    └── fault_type_B/
        └── fault1.csv

Sample file (normal/run1.csv):

accel_x,accel_y,accel_z
0.012,0.005,-0.982
0.015,0.008,-0.979
0.010,0.003,-0.985
...

Config:

dataset:
  dataset_name: 'vibration_data'
  input_data_path: '/data/vibration_dataset'

data_processing_feature_extraction:
  feature_extraction_name: 'Generic_256Input_FFTBIN_16Feature_8Frame'
  variables: 3    # x, y, z axes

5.1.8. Class Balancing

For best results, try to have similar sample counts per class.

If classes are imbalanced:

Collect more data for minority classes
Use data augmentation (gain variation)
Adjust training parameters

data_processing_feature_extraction:
  gain_variations: {fault_type_A: [0.9, 1.1]}  # Augment minority class

5.1.9. Common Issues

“Dimension mismatch” error

All files must have the same number of columns. Check for:

Extra header rows
Missing columns in some files
Different delimiters

“Empty file” error

Ensure files contain actual data, not just headers.

Class not detected

Check folder names don’t contain special characters
Ensure files exist in class folders
Verify file extensions are supported

5.1.10. Key Differences: Classification vs Regression vs Forecasting

The following table summarizes the key structural and behavioral differences across the three time series task types.

Aspect	Classification	Regression	Forecasting
Folder Structure	`classes/{class_name}/`	`files/`	`files/`
Target Location	Folder name (implicit)	Last column of each file	Specified via `target_variables`
Target Type	Discrete class label	Continuous value (averaged over window)	Future value at `forecast_horizon` steps
Annotations	Optional (auto-generated)	Required	Required
Feature Extraction	Supported (FFT, wavelets, etc.)	Supported (`SimpleWindow` required)	Not supported (raw time series only)
Loss Function	CrossEntropyLoss	MSELoss	HuberLoss
Evaluation Metrics	Accuracy, F1-score	MSE, R-squared	SMAPE, R-squared
`output_int`	Can be True	Can be True	Must be False
`SimpleWindow`	Optional	Mandatory	Mandatory
`forecast_horizon`	N/A	N/A	Required parameter