search

open-mmlab / mmdetection

OpenMMLab Detection Toolbox and Benchmark
https://mmdetection.readthedocs.io
Apache License 2.0
29.21k stars 9.4k forks source link

Working Swin Transformer Demo with Custom Dataset #11041

Open ozan956 opened 11 months ago

ozan956 commented 11 months ago

Greetings to everyone,

I have been trying to train with my own dataset using Swin transformer models for a while. However, I have not been successful in any way. If anyone has done this before, can you share the Google Collab link or github repo? I really need this. It would be much better if you can share it with the dataset because I think I might be having a problem with my dataset format.

hhaAndroid commented 11 months ago

@ozan956 You should provide the error message. If you are not familiar with customizing datasets, you can refer to configs/faster_rcnn/faster-rcnn_r50-caffe_fpn_ms-1x_coco-person-bicycle-car.py . Just pay attention to the meainfo part.

ozan956 commented 11 months ago

Thank you for your message. I only wanted to use the object detection part of the Swin Transformer without instance segmentation. I learned how to do this through the following thread: https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/issues/50 . If a beginner of mmdetection, like myself, follows a tutorial and attempts to reimplement it using Swin Transformer models, they may encounter an error because there is no mask annotation in the Kitty_tiny dataset (as discussed in https://github.com/open-mmlab/mmdetection/issues/8058 ). Therefore, if the relevant lines in the configuration files are adjusted, the problem can be resolved. I plan to share my Colab notebook once I achieve a stable result.

ozan956 commented 11 months ago

Just one more question @hhaAndroid , will swin transformer v2 models be added ?