Skip to content

Dataset Setup & Checkpoint Size Issue in MMSegmentation #3850

Open
@hotelbread

Description

@hotelbread

Hi,

Thanks to your hard work, I’m honored to be able to start segmentation so easily. As part of my initial learning, I went through the MMSegmentation tutorial to fine-tune the SegFormer model for my specific task, and I encountered a few questions.

First, I have an open dataset where each sample consists of a pair of images: a regular image and its corresponding segmentation mask. However, I do not have any information about the color values or the number of classes in the dataset—I only have the image pairs. Given this situation, how can I properly prepare this dataset for training? I attempted to create a custom dataset class, but I’m unsure how to define the label and palette information. How should I handle this?

Second, I have a question about the .pth files saved during training. As shown in the attached image, I monitored a full training cycle without making any specific changes to the configuration. I noticed that the .pth model files increase in size as the iterations progress. Generally, model files of the same architecture should have similar sizes, so why is this happening?

Image

If you need any additional details to answer my questions, please feel free to ask! I’ll be happy to share everything I’ve tried so far.

Thanks! 😊

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions