Any advice on adding a CNN backbone？

[JolyneAnasui]

i want to add a CNN like ResNet50 or VGG as backbone, which output feature map as the input of swin-transformer
cause my own dataset only include a few thousands images, so i don't want to take the raw image or just normalized image as the input of swin-transformer
Any advice on adding a CNN backbone, requirement of the input size of swin-transformer seems to be strict.

[zeliu98]

Hi, for custom input size you can simply add some paddings, our detection and segmentation repo give some example:

• https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/blob/6a979e2164e3fb0de0ca2546545013a4d71b2f7d/mmdet/models/backbones/swin_transformer.py#L433-L436
• https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/blob/6a979e2164e3fb0de0ca2546545013a4d71b2f7d/mmdet/models/backbones/swin_transformer.py#L284-L286
• https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/blob/6a979e2164e3fb0de0ca2546545013a4d71b2f7d/mmdet/models/backbones/swin_transformer.py#L215-L218
• ...

[ancientmooner]

When you have only a few thousands images, one way to leverage Swin Transformer is to do fine-tuning from the pre-trained models.

[wutongshu819274135]

i want to add a CNN like ResNet50 or VGG as backbone, which output feature map as the input of swin-transformer cause my own dataset only include a few thousands images, so i don't want to take the raw image or just normalized image as the input of swin-transformer Any advice on adding a CNN backbone, requirement of the input size of swin-transformer seems to be strict.

hello，I also use a small data set. Can I exchange it。qq：819274135