Validation image size differs from training size within each epoch

[SPONGEBOB-du]

Hi, I'm training a YOLOv8 model using the imgsz=320 argument. I expected both training and validation images to be resized to 320x320. However, during training, I noticed that the image sizes during validation differ from the training image size within each epoch.

Is there a way to force validation to always use the same size (320x320) as training?
Also, I am not using multi-scale training or custom dataloaders.

Here's the code I'm using:

results = model.train(data='coco128.yaml', epochs=2, batch=128, device=[0,1], # 指定GPU索引列表 workers=0, imgsz=320, )

[UltralyticsAssistant]

👋 Hello @SPONGEBOB-du, thank you for your interest in Ultralytics 🚀! We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Join the Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

---

This is an automated response to help you get started quickly. An Ultralytics engineer will also assist you here soon 😊

[glenn-jocher]

Hi @SPONGEBOB-du,

This behavior occurs because by default, validation uses rectangular batching (rect=True) while training uses square images. The rectangular batching preserves aspect ratios during validation to improve efficiency and accuracy.

To force validation to use exactly 320×320 like your training, add the rect=False parameter:

results = model.train(
    data='coco128.yaml', 
    epochs=2, 
    batch=128,
    device=[0,1],
    workers=0,
    imgsz=320,
    rect=False  # Force square images during validation
)

This will ensure both training and validation use the same 320×320 square image size throughout the process.

[SPONGEBOB-du]

I'm sorry, but I’ve already tried setting rect=False. However, during training, the validation image size still changes in each epoch. For example, I passed in images with shape [128, 3, 320, 320], but during validation, the input shape changes to [128, 3, 352, 352].

Is there a way to enforce a fixed validation image size throughout training?

[Y-T-G]

You can change rect to False here

ultralytics/ultralytics/models/yolo/detect/train.py

Line 65 in 5e892b6

return build_yolo_dataset(self.args, img_path, batch, self.data, mode=mode, rect=mode == "val", stride=gs)

[SPONGEBOB-du]

thanks so much. it works.

[glenn-jocher]

You're welcome! I'm glad that resolved the issue for you. The behavior you experienced is actually by design - during training, the validator defaults to rect=True for efficiency, but modifying that line in the source code gives you the exact control you needed over validation image sizes.