yehao
yehao
It used finetune.
In my test, mobilenet-v2-lite is much slower than mobilenet-v1 in desktop GPU(Google say it is normal) hardware: 1070ti + cuda9 + cudnn v7 mobilenet v1: 181 fps mobilenet v2: 100...
you can see Table 1 or Tabel2, it exists activate after pool
You can train it multiple times.
@OMG59E the origin SSD has 38x38 feature map.
@gombru It maybe your dataset has 1 or 4 channel images.
@kuan-wang Hi, what are your mobilenet v3 small 1.0 params? I calculate it and get 3.09 million (not include BN params), the paper claims 2.9 Million.
The resolution size can affect mAP, if the resolution size is bigger, the mAP is higher. How do you get mAP 0.74? the author just trains a model and its...
@CrazyAlan padding is different between caffe and tensorflow.
It may be a mistake.