Imagenet pretrained weights

[baoruxiao]

Hi Xiaochus,

Would you mind sharing your imagenet pretrained weights?

[JonathanCMitchell]

Hi yes would you mind sharing them? I could really use them right now. Thanks!

[gustavz]

Same here! Would be very useful!

[xiaochus]

Unfortunately, due to limited computational power, I only experimented on the cifar dataset and did not use imagenet.

[RadEdje]

Hi, I was hoping to retrain this mobileNetv2 with imagenet weights. So based on this thread it hasn't been trained on that correct? So i have to go the the main mobileNetV2.py from keras team at this github repo:

https://github.com/keras-team/keras/blob/master/keras/applications/mobilenetv2.py

and use that for re-training? thanks.

[alyato]

@xiaochus
i m curious about these lines.

x = GlobalAveragePooling2D()(x)
x = Reshape((1, 1, 1280))(x)
x = Dropout(0.3, name='Dropout')(x)
x = Conv2D(k, (1, 1), padding='same')(x)
x = Activation('softmax', name='softmax')(x)
output = Reshape((k,))(x)

why it is this?
not

x = GlobalAveragePooling2D()(x)
x = Dense(k)(x)
x = Activation('softmax', name='softmax')(x)

Do some guys give one explain?
thanks.