Questions about train and validation

[kennedyk1]

Hello everyone!

I have a few questions about the use of train, test, and validation sets during my model training process that remain unclear. I’ll describe my situation below to provide more context:

• I split my dataset into two folders: train and test.
• In the rgb.yaml file, I set the test folder as the validation set.

path: '/home/user/my_datasets/'
train: 'dataset_1/train/rgb/'
val: 'dataset_1/test/rgb/'
# class names
names:
  0: 'human'

• In the main code, I disabled early stopping by setting patience=0 and trained the model for 50 epochs.
• I used the low-scratch.yaml hyperparameter profile to enhance data augmentation.

My main code:

from ultralytics import YOLO
datasets = ['rgb.yaml']
yolo_models = ['yolov6s.yaml','yolov5s.yaml','yolov8s.yaml','yolov9s.yaml','yolov10s.yaml','yolo11s.yaml']
epochs = 50

for dataset in datasets:
    for yolo_model in yolo_models:
        model = YOLO(yolo_model,task='detect') # Create a new model to train
        model.train(
            epochs = epochs, #NUMBER EPOCHS TO TRAIN
            batch = 0.9, #TO USE 90% GPU MEMORY
            single_cls = True, #TO TRAIN WITH ONLY ONE CLASS
            imgsz = 640, # TO REZISE IMAGES, DEFAULT 640
            data = dataset, #TO DEFINE YAML DATASET FILE
            patience = 0, #TO AVOID EARLY STOP
            save = True, #TO SAVE CHECKPOINTS AND FINAL MODEL WEIGHTS
            project = yolo_model.split('.')[0], #NAME OF PROJECT
            name = dataset.split('.')[0], #SUB-NAME OF PROJECT or MODALITY
            plots = True #TO SHOW PLOTS OF TRAINNING AND VALIDATION METRICS
        )

Based on the details above, I have the following questions:

• When is data augmentation applied?
• Is data augmentation performed on the train set, the validation set, or both?
• At what stage is validation conducted?
• Can my test (used as validation) folder be considered "unseen" data, given that I have disabled early stopping?
• Can I consider given metrics (P, R, mAP50 and mAP50:95) at end training as test evaluation, given that I have disabled early stopping?