search

xiaolai-sqlai / mobilenetv3

mobilenetv3 with pytorch，provide pre-train model
MIT License
1.6k stars 340 forks source link

Difference on SE from paper #15

Open triangleCZH opened 5 years ago

triangleCZH commented 5 years ago

1) I am just curious, why do you add batch normalization inside SeModule? Is there a reference to do so? 2) Please correct me if I make a mistake: I think SeModule should be added between dw and pw-linear, but your code seems to add that after pw-linear and right before residual connection 3) Do you think it's necessary to consider expand_ratio = 1? When expand_channel == output_channel, I feel that pw might be redundant, since the shape won't change a bit after pw.

Thank you!

ujsyehao commented 4 years ago
  1. In efficientnet, there is NO BN in SE.
  2. SE should be added between 3x3 dw and 1x1 pw.
  3. it is redundant. You can refer to mobilenet v2.