keras-team/keras-applications

add resnet18 and resnet34

dulucas opened this issue · 11 comments

dulucas commented

Hi! could you please add resnet18 and resnet34 with their pre-trained weights on Imagenet? Thanks!

jjedele commented

I also think this would be a good addition - Resnet50 can be a bit heavy for some tasks. Are there any reasons for them not being included? It seems fairly straight-forward to add these configurations to the existing codebase.

jjedele commented

Also because there is another Resnet implementation in Keras Contrib including 18 and 34.

  • 👍2
WurmD commented

Some feedback here about for reasons it's not included would be appreciated.
Also, keras-contrib says "is deprecated. Use TensorFlow Addons.", but the https://github.com/keras-team/keras-contrib/tree/master/keras_contrib/applications folder which includes resnet18 is not currently in https://github.com/tensorflow/addons/tree/master/tensorflow_addons

dulucas commented

Hi all, I use resnet model provided in this repo instead and it works fine for me, https://github.com/qubvel/classification_models

  • 👍10
Linriversluv commented

Yeah, it would be great if they're included.

hamddan4 commented

Also agree. If you have all the blocks for ResNet50, ResNet101, ResNet152, why not adding a basic structure as keras_contrib does so you can create your own ResNet?

  • 👍2
zaccharieramzi commented

For anyone looking for a maintained official implementation, it is available in tensorflow/models.
There is however a slight "bug" in the implementation because it has an extra convolution compared to the paper and to PyTorch, but I created a PR to fix it.

zaccharieramzi commented

The model is now in keras-cv after this PR.