Tokio

mobilenetv3 代码实现似乎和论文有出入

1

Mobilenetv3模块实现部分好像少了一个pointwise convolution, 见 https://github.com/d-li14/mobilenetv3.pytorch/blob/master/mobilenetv3.py#:~:text=and%20inp%20%3D%3D%20oup-,if%20inp%20%3D%3D%20hidden_dim%3A,),-else%3A 在line94-104 这部分只适用于第一个模块，代码可以继续简化

I found that the MobileNetv2 has more shortcut with conv1d and batchnorm ops when stride==1 compared to original paper, which leads to the high performance of the modified model. I...

关于channel-wise convolution

1

您好，channel-wise convolution的实现方式是utils/ops.py文件里面的rev_conv2d函数么？ 如果是这样，似乎实现方式有点不对。conv1d应该在[-1, num_input_channels x h x w,1]形状的数据上做卷积，而不是在[-1, num_input_channels，h x w]上做卷积。

cannot import name 'COCODataset' from 'yolox.data' when running