Chengpeng Chen

Hi, would you update the training setting for each dataset? I can't reproduce the result yet...

Ghost-ResNet-50 结构flops和parameters

9

我把ResNet-50 的conv（除了第一个7x7xconv和最后的fc）替换为ghost module，得到的模型（下图是对应的onnx结构）： 1. Ghost-ResNet-50 (s=2), flops=2.15G, params=13.95M 2. Ghost-ResNet-50 (s=4), flops=1.18G, params=8.15M 其中两个模型的flops跟paper中的对得上，但是params 都对不上，请问可以提供Ghost-ResNet-50的代码吗？

In the original RCNN, cfg.TRAIN.BATCH_ROIS rois are selected. Does this implementation use the same setting and just modify the cls loss? Or, much more Rois are selected, following the spirit...

Will you release MobileNetV4 checkpoints?

5

I can only find \'MobileNetV4-Conv-Large-fp32.tflite\' in other repos, but it is only for inference. Do you plan to release the checkpoints, including the un-merged BN parameters. Thanks!

Support RepGhostNetV2 conv s/m/l, based on MobileNetV4, without weights though

4

Build RepGhostNetV2 based on MobileNetV4. Our RepGhostNetV2 differs to MobileNetV4 only in the shortcut connections. As an alternative efficient backbone to MobileNetV4, RepGhostNetV2 has exactly the same parameters and FLOPs...