Choosing imgsz for training on Custom dataset with fixed-size Non-square images

[assafzam]

Hi,
I have a dataset with images in shape (1350, 200, 3).
Assuming that:

• best precision and recall are the main goals
• resources are not limited

How would you suggest setting the imgsz parameter at training?
(It can be only an integer and not a tuple)

[Laughing-q]

@assafzam From your shape 1350x200, assuming it'll lose objects if resize your images to a lower resolution I recommend you to set imgsz=1376 and rect=True, imgsz=1376 is for keeping your original image size and also satisfy divided by 32, and rect=True is for rectangle training which actually would train with 1376x224 and save a lot memory compare to the square size 1376x1376.

[glenn-jocher]

@silvada95 it's up to you to experiment with the best training settings for your custom data.

[silvada95]

In this case, what is the effect of the parameter "Stride" on such rectangular shapes? Do you recommend changing the "256" parameter used to build the strides?

[glenn-jocher]

Stride is a model property, it's not something you set. P5 models have output strides of 8, 16, 32.

[Rabie-el-hakouni]

i have a dataset of images with different resolution when i use the pretrained model YOLOv8n classifier , i use imgsz = 224 , that mean the images reshaped to 224*224 ?

[glenn-jocher]

@Rabie-el-hakouni yes, setting imgsz=224 will resize your images to 224x224 during training and inference, which may distort non-square images if they are not padded appropriately. YOLOv8 will handle the resizing and padding automatically to maintain the aspect ratio if --rect is also used during training.

[Vishwajeet-assert]

@glenn-jocher
For example, you can set imgsz=(3840, 2160) to resize your input images to 3840x2160. does not work kindly explain how to use w,h format.

[Totoro-wen]

@Vishwajeet-assert
I also met the same problem. Have you solved it? I tried to modify rect=True in default.yaml, but it still didn't work

[4alexw]

Have you tried making the height a multiple of 32 like 3840x2176? I don't know if it explicitly says in the documentation anywhere but I do remember reading somewhere that during training images need to have both sides equal to a multiple of 32 because that is the maximum stride length in the backbone of the model. Let me know if this helps because I am probably going to need to test this out at some point myself in the near future. Also if someone else knows more about this then elaborate on this because I am a little confused about how the image sizing is supposed to work myself.

[glenn-jocher]

@4alexw hello everyone,

When training YOLOv8 on a custom dataset with non-square images, it's important to choose an imgsz that respects the aspect ratio of your dataset while also being a multiple of the maximum stride of the network (typically 32 or 64). Since your images are of shape (1350, 200, 3), you'll want to resize them in a way that maintains the aspect ratio as closely as possible.

Given that resources are not a constraint for you, and you're aiming for the best precision and recall, you should select an imgsz that is a multiple of 32 and maintains the aspect ratio of your original images. For instance, you could scale your width up to the nearest multiple of 32 that maintains the aspect ratio with the height. However, since imgsz can only be an integer, you would choose a size that scales both dimensions while keeping the aspect ratio intact.

For your specific case, you could calculate the target size by finding the nearest multiple of 32 that maintains the aspect ratio. Remember that during training, the images will be resized to a square format, so you should choose a size that minimizes the distortion.

Please note that setting rect=True in the configuration file allows for rectangular training, which can help maintain the aspect ratio of your images. However, this feature is not available in all versions of YOLO, and you should refer to the documentation for specific instructions on how to enable and use this feature in YOLOv8.

For further details on how to configure your training setup, please refer to the documentation at https://docs.ultralytics.com.

[Vishwajeet-assert]

no

…

On Tue, 11 Apr 2023 at 15:58, chenewy ***@***.***> wrote: @Vishwajeet-assert <https://github.com/Vishwajeet-assert> I also met the same problem. Have you solved it?I tried to modify rect=true in default.yaml, but it still didn't work — Reply to this email directly, view it on GitHub <#1100 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A6L37UI2JIZCDIRLBOGDPSDXAUW37ANCNFSM6AAAAAAVEELVJA> . You are receiving this because you were mentioned.Message ID: ***@***.*** com>

-- Vishwajeet Singh Senior AI Engineer [image: mobile-icon] 9027570172 [image: phone-icon] ***@***.*** [image: mobile-icon] www.assertai.com [image: mobile-icon] Assert SecureTech Pvt. Ltd, C-605/606, 6th Floor, Kailash Business Park, Park Site, Vikhroli West, Mumbai 400079 [image: linkedin icon] <https://linkedin.com/in/vishwajeet-singh-66aab2155> <https://www.hubspot.com/email-signature-generator?utm_source=create-signature> <https://www.hubspot.com/email-signature-generator?utm_source=create-signature>

[glenn-jocher]

@Vishwajeet-assert hello,

For training on a custom dataset with non-square images, such as your (1350, 200, 3) images, you should set the imgsz parameter to the largest dimension of your dataset images, which in your case is 1350. YOLOv8 will automatically handle the aspect ratio and resize your images accordingly during training while maintaining the original aspect ratio.

Since resources are not a constraint for you, using the largest dimension will allow the model to train on the highest resolution possible, which is beneficial for achieving the best precision and recall.

Remember to monitor your GPU memory usage to ensure that it does not exceed the available resources, as very large image sizes can lead to out-of-memory errors.

Happy training! 😊🚀