search

megvii-model / YOLOF

MIT License
809 stars 115 forks source link

Weight Initialization method in DilatedEncoder #31

Closed developer0hye closed 3 years ago

developer0hye commented 3 years ago

https://github.com/megvii-model/YOLOF/blob/6189487b80601dfeda89302c22abac060f977785/playground/detection/coco/yolof/yolof_base/encoder.py#L64-L78

Why the weight initialization method for lateral_conv and fpn_conv is different from the weight initialization method for conv layer in dilated_encoder_blocks?

chensnathan commented 3 years ago

For lateral_conv and fpn_conv, we follow the weight initialization method in FPN. For others, we follow the weight initialization method in the head of RetinaNet.

We did not tune the methods of weight initialization. You can try other methods.

developer0hye commented 3 years ago

@chensnathan Thanks for your reply!