K-Fold Cross Validation with Ultralytics

Introduction

This comprehensive guide illustrates the implementation of K-Fold Cross Validation for object detection datasets within the Ultralytics ecosystem. We'll leverage the YOLO detection format and key Python libraries such as sklearn, pandas, and PyYaml to guide you through the necessary setup, the process of generating feature vectors, and the execution of a K-Fold dataset split.

Whether your project involves the Fruit Detection dataset or a custom data source, this tutorial aims to help you comprehend and apply K-Fold Cross Validation to bolster the reliability and robustness of your machine learning models. While we're applying k=5 folds for this tutorial, keep in mind that the optimal number of folds can vary depending on your dataset and the specifics of your project.

Without further ado, let's dive in!

Setup

• Your annotations should be in the YOLO detection format.

• This guide assumes that annotation files are locally available.

• For our demonstration, we use the Fruit Detection dataset.

  • This dataset contains a total of 8479 images.
  • It includes 6 class labels, each with its total instance counts listed below.

• Necessary Python packages include:

  • ultralytics
  • sklearn
  • pandas
  • pyyaml

• This tutorial operates with k=5 folds. However, you should determine the best number of folds for your specific dataset.

• Initiate a new Python virtual environment (venv) for your project and activate it. Use pip (or your preferred package manager) to install:

  • The Ultralytics library: pip install -U ultralytics. Alternatively, you can clone the official repo.
  • Scikit-learn, pandas, and PyYAML: pip install -U scikit-learn pandas pyyaml.

• Verify that your annotations are in the YOLO detection format.

  • For this tutorial, all annotation files are found in the Fruit-Detection/labels directory.

Generating Feature Vectors for Object Detection Dataset

1. Start by creating a new Python file and import the required libraries.

   import datetime
   import shutil
   from pathlib import Path
   from collections import Counter
   
   import yaml
   import numpy as np
   import pandas as pd
   from ultralytics import YOLO
   from sklearn.model_selection import KFold
   

2. Proceed to retrieve all label files for your dataset.

   dataset_path = Path('./Fruit-detection') # replace with 'path/to/dataset' for your custom data
   labels = sorted(dataset_path.rglob("*labels/*.txt")) # all data in 'labels'
   

3. Now, read the contents of the dataset YAML file and extract the indices of the class labels.

   yaml_file = 'path/to/data.yaml'  # your data YAML with data directories and names dictionary
   with open(yaml_file, 'r', encoding="utf8") as y:
       classes = yaml.safe_load(y)['names']
   cls_idx = sorted(classes.keys())
   

4. Initialize an empty pandas DataFrame.

   indx = [l.stem for l in labels] # uses base filename as ID (no extension)
   labels_df = pd.DataFrame([], columns=cls_idx, index=indx)
   

5. Count the instances of each class-label present in the annotation files.

   for label in labels:
       lbl_counter = Counter()
   
       with open(label,'r') as lf:
           lines = lf.readlines()
   
       for l in lines:
           # classes for YOLO label uses integer at first position of each line
           lbl_counter[int(l.split(' ')[0])] += 1
   
       labels_df.loc[label.stem] = lbl_counter
   
   labels_df = labels_df.fillna(0.0) # replace `nan` values with `0.0`
   

6. The following is a sample view of the populated DataFrame:

                                                          0    1    2    3    4    5
   '0000a16e4b057580_jpg.rf.00ab48988370f64f5ca8ea4...'  0.0  0.0  0.0  0.0  0.0  7.0
   '0000a16e4b057580_jpg.rf.7e6dce029fb67f01eb19aa7...'  0.0  0.0  0.0  0.0  0.0  7.0
   '0000a16e4b057580_jpg.rf.bc4d31cdcbe229dd022957a...'  0.0  0.0  0.0  0.0  0.0  7.0
   '00020ebf74c4881c_jpg.rf.508192a0a97aa6c4a3b6882...'  0.0  0.0  0.0  1.0  0.0  0.0
   '00020ebf74c4881c_jpg.rf.5af192a2254c8ecc4188a25...'  0.0  0.0  0.0  1.0  0.0  0.0
    ...                                                  ...  ...  ...  ...  ...  ...
   'ff4cd45896de38be_jpg.rf.c4b5e967ca10c7ced3b9e97...'  0.0  0.0  0.0  0.0  0.0  2.0
   'ff4cd45896de38be_jpg.rf.ea4c1d37d2884b3e3cbce08...'  0.0  0.0  0.0  0.0  0.0  2.0
   'ff5fd9c3c624b7dc_jpg.rf.bb519feaa36fc4bf630a033...'  1.0  0.0  0.0  0.0  0.0  0.0
   'ff5fd9c3c624b7dc_jpg.rf.f0751c9c3aa4519ea3c9d6a...'  1.0  0.0  0.0  0.0  0.0  0.0
   'fffe28b31f2a70d4_jpg.rf.7ea16bd637ba0711c53b540...'  0.0  6.0  0.0  0.0  0.0  0.0
   

The rows index the label files, each corresponding to an image in your dataset, and the columns correspond to your class-label indices. Each row represents a pseudo feature-vector, with the count of each class-label present in your dataset. This data structure enables the application of K-Fold Cross Validation to an object detection dataset.

K-Fold Dataset Split

1. Now we will use the KFold class from sklearn.model_selection to generate k splits of the dataset.

   • Important:
     • Setting shuffle=True ensures a randomized distribution of classes in your splits.
     • By setting random_state=M where M is a chosen integer, you can obtain repeatable results.

   ksplit = 5
   kf = KFold(n_splits=ksplit, shuffle=True, random_state=20)   # setting random_state for repeatable results
   
   kfolds = list(kf.split(labels_df))
   

2. The dataset has now been split into k folds, each having a list of train and val indices. We will construct a DataFrame to display these results more clearly.

   folds = [f'split_{n}' for n in range(1, ksplit + 1)]
   folds_df = pd.DataFrame(index=indx, columns=folds)
   
   for idx, (train, val) in enumerate(kfolds, start=1):
       folds_df[f'split_{idx}'].loc[labels_df.iloc[train].index] = 'train'
       folds_df[f'split_{idx}'].loc[labels_df.iloc[val].index] = 'val'
   

3. Now we will calculate the distribution of class labels for each fold as a ratio of the classes present in val to those present in train.

   fold_lbl_distrb = pd.DataFrame(index=folds, columns=cls_idx)
   
   for n, (train_indices, val_indices) in enumerate(kfolds, start=1):
       train_totals = labels_df.iloc[train_indices].sum()
       val_totals = labels_df.iloc[val_indices].sum()
   
       # To avoid division by zero, we add a small value (1E-7) to the denominator
       ratio = val_totals / (train_totals + 1E-7)
       fold_lbl_distrb.loc[f'split_{n}'] = ratio
   

The ideal scenario is for all class ratios to be reasonably similar for each split and across classes. This, however, will be subject to the specifics of your dataset.

1. Next, we create the directories and dataset YAML files for each split.

   supported_extensions = ['.jpg', '.jpeg', '.png']
   
   # Initialize an empty list to store image file paths
   images = []
   
   # Loop through supported extensions and gather image files
   for ext in supported_extensions:
       images.extend(sorted((dataset_path / 'images').rglob(f"*{ext}")))
   
   # Create the necessary directories and dataset YAML files (unchanged)
   save_path = Path(dataset_path / f'{datetime.date.today().isoformat()}_{ksplit}-Fold_Cross-val')
   save_path.mkdir(parents=True, exist_ok=True)
   ds_yamls = []
   
   for split in folds_df.columns:
       # Create directories
       split_dir = save_path / split
       split_dir.mkdir(parents=True, exist_ok=True)
       (split_dir / 'train' / 'images').mkdir(parents=True, exist_ok=True)
       (split_dir / 'train' / 'labels').mkdir(parents=True, exist_ok=True)
       (split_dir / 'val' / 'images').mkdir(parents=True, exist_ok=True)
       (split_dir / 'val' / 'labels').mkdir(parents=True, exist_ok=True)
   
       # Create dataset YAML files
       dataset_yaml = split_dir / f'{split}_dataset.yaml'
       ds_yamls.append(dataset_yaml)
   
       with open(dataset_yaml, 'w') as ds_y:
           yaml.safe_dump({
               'path': split_dir.as_posix(),
               'train': 'train',
               'val': 'val',
               'names': classes
           }, ds_y)
   

2. Lastly, copy images and labels into the respective directory ('train' or 'val') for each split.

   • NOTE: The time required for this portion of the code will vary based on the size of your dataset and your system hardware.

   for image, label in zip(images, labels):
       for split, k_split in folds_df.loc[image.stem].items():
           # Destination directory
           img_to_path = save_path / split / k_split / 'images'
           lbl_to_path = save_path / split / k_split / 'labels'
   
           # Copy image and label files to new directory (SamefileError if file already exists)
           shutil.copy(image, img_to_path / image.name)
           shutil.copy(label, lbl_to_path / label.name)
   

Save Records (Optional)

Optionally, you can save the records of the K-Fold split and label distribution DataFrames as CSV files for future reference.

folds_df.to_csv(save_path / "kfold_datasplit.csv")
fold_lbl_distrb.to_csv(save_path / "kfold_label_distribution.csv")

Train YOLO using K-Fold Data Splits

1. First, load the YOLO model.

   weights_path = 'path/to/weights.pt'
   model = YOLO(weights_path, task='detect')
   

2. Next, iterate over the dataset YAML files to run training. The results will be saved to a directory specified by the project and name arguments. By default, this directory is 'exp/runs#' where # is an integer index.

   results = {}
   
   # Define your additional arguments here
   batch = 16
   project = 'kfold_demo'
   epochs = 100
   
   for k in range(ksplit):
       dataset_yaml = ds_yamls[k]
       model.train(data=dataset_yaml,epochs=epochs, batch=batch, project=project)  # include any train arguments
       results[k] = model.metrics  # save output metrics for further analysis
   

Conclusion

In this guide, we have explored the process of using K-Fold cross-validation for training the YOLO object detection model. We learned how to split our dataset into K partitions, ensuring a balanced class distribution across the different folds.

We also explored the procedure for creating report DataFrames to visualize the data splits and label distributions across these splits, providing us a clear insight into the structure of our training and validation sets.

Optionally, we saved our records for future reference, which could be particularly useful in large-scale projects or when troubleshooting model performance.

Finally, we implemented the actual model training using each split in a loop, saving our training results for further analysis and comparison.

This technique of K-Fold cross-validation is a robust way of making the most out of your available data, and it helps to ensure that your model performance is reliable and consistent across different data subsets. This results in a more generalizable and reliable model that is less likely to overfit to specific data patterns.

Remember that although we used YOLO in this guide, these steps are mostly transferable to other machine learning models. Understanding these steps allows you to apply cross-validation effectively in your own machine learning projects. Happy coding!

Comments