训练时loss能不能降下来？Can the loss be reduced in your project?

PINTO0309 / MobileNetv2-SSDLite

My proprietary procedure. Caffe implementation of SSD and SSDLite detection on MobileNetv2, converted from tensorflow.

MIT License

23 stars 11 forks source link

Closed yulong112 closed 5 years ago

yulong112 commented 5 years ago

The loss would not decrease below 5 when I use "Chuanqi305/MobileNetv2-SSDLite" project. Can the loss be reduced in your project?

PINTO0309 commented 5 years ago

@yulong112

Can the loss be reduced in your project?

No. Because Pure Caffe does not support Relu6. I think that if you have custom Caffe corresponding to Relu6 you can reduce the loss.

PINTO0309 commented 5 years ago

@yulong112 Custom Caffe with ReLU6 enabled is below. https://github.com/chuanqi305/ssd.git