why did your ResNet-50 improve 2X speed? about channel-pruning HOT 7 CLOSED

1. A common practice is to merge batchnorm into conv when inference
2. ResNet has only 1 pooling layer.
3. 2X is in terms of FLOPs, not GPU real time

from channel-pruning.

@yihui-he
1.merge batchnorm into conv when inference
--answer: The maximum increase was 10% in speed.
2.2X is in terms of FLOPs, not GPU real time,
--problem: how about your paper to practice?

from channel-pruning.

We only compared GPU acceleration on VGG-16.

from channel-pruning.

So it was not suit to ResNet to prune.

from channel-pruning.

from channel-pruning.

yes, It should be full rank.

from channel-pruning.

@eeric Hi,you said that "yes, It should be full rank.",do you mean the rankdic in code should be set as the same as the output channel of the resnet? Or how should I change it for resnet? Thank you so much!
In the author's the code for vgg the rankdic is as follows:
rankdic = {'conv1_1': 17,
'conv1_2': 17,
'conv2_1': 37,
'conv2_2': 47,
'conv3_1': 83,
'conv3_2': 89,
'conv3_3': 106,
'conv4_1': 175,
'conv4_2': 192,
'conv4_3': 227,
'conv5_1': 398,
'conv5_2': 390,
'conv5_3': 379}

from channel-pruning.