mobilenet performs worse than inceptionv1 on size and speed about mobilenet HOT 8 CLOSED

A known bug:
slim.separable_convolution2d used mistakenly

I am working on this.

from mobilenet.

I have used tf.nn.depthwise_conv2d instead. The model size should be right now.

from mobilenet.

I set num_outputs=None in the separable_convolution2d. Does it work the same way as depthwise_conv2d?

from mobilenet.

Look good! Thanks.
I will train the first model this weekend.

from mobilenet.

@Zehaos Would it be the same if you replaced your function for depthwise separable conv2d with the official function from tensorflow in https://github.com/tensorflow/tensorflow/blob/08cb0ba5431ef2508c6ca63c498023a84f7b4a18/tensorflow/python/layers/convolutional.py#L870 ?

It includes the depth_multiplier in the function as well, which could make the code more concise. Did you see any difference from implementing this official function instead?

from mobilenet.

@kwotsin According to tf doc,I think 'tf.nn.separable_conv2d()' combine depthwise conv and pointwise conv together, whithout bn nor relu between them, that is the problem.

from mobilenet.

@Zehaos Hmm do you think it is possible to use slim.batch_norm as the normalizer_fn argument in slim.conv2d or slim.separable_convolution2d to use the batch norm instead? Relu is included by default in those 2 slim functions, and go by conv-->batch_norm-->relu, so I think it might work.

from mobilenet.

@kwotsin yes, you are right. tf is so fexiable.

from mobilenet.