xiaochus/MobileNetV3

How to change depth multiplier values?

anilsathyan7 opened this issue · 2 comments

anilsathyan7 commented

Hey in you code depth multiplier value is 1 and width multiplier alpha can be configured and theoretically both seems to be different.If i change the depth multiplier value as 0.5 in depthwise conv2d(like waht appears in official mobilenetv2 or deeplab) it shows error. If i want to decrease the depth multiplier how can it be done?

Ref

  1. keras-team/keras#10349
  2. https://github.com/keras-team/keras-applications/blob/master/keras_applications/mobilenet.py
henry2304 commented

Hi @anilsathyan7, I was also trying to get the alpha parameter to work, and set the alpha parameter to 0.5 by modifying line 106 of train_cls.py:
model = MobileNetV3_Small(shape, n_class, 0.5).build()
I don't seem to get any error, but the models (.h5 weights) I obtain still have size 12MB, which was the same as alpha=1.0. Did you have any luck on figuring this out, Anil? In case you or @xiaochus have any suggestions, please let me know. Thanks for helping implement this code, @xiaochus!

anilsathyan7 commented

Hi @henry2304 have you tried official implementation from keras-applications?
Check out: keras-team/keras-applications#183