Where can I download the pretained model based on Imagenet? about keras-resnext

[titu1994]

ResNeXt is not supported by Keras or Tensorflow. Therefore there are no weights.

[John1231983]

Hi. Because keras does not support group convolution yet. So how did you solve with it in resnext? Does it loss some performance without using it?

[titu1994]

I used lambda layers to slice off filters and word on them separately and then concatenated that. However, this is a low and mostly wrong implementation with none of the speed benefits of grouped convolutions.

[John1231983]

I see. So using your way just makes the code run slowly, and it will not affect the accuracy. Am I right?

[titu1994]

Theoretically, it may affect accuracy as well. And it's a lot slower, so I recommend not to use it.

There are better models out there such as Dense and SparseNets, SENets and NASNet.

[John1231983]

SENets also use ResneXt :(. I am finding the Se-resneXT implementation

[titu1994]

The squeeze and excitation mechanism is generic to any architecture. There are se variants to ordinary ResNets, DenseNets, Inception and other types of nets in my repo.

Coincidentally,vi would not advice to use the SE-ResNext in my repo either