Here is my implementation of residual networks on Keras (on Theano). I’ll update readme.md as there’s update more up-to-date information would be there.

**14 Apr 2016.** I updated the implementation. Now the example.py working properly, showing reasonable performances, and re-written with Keras 1.0 API (which is AWESOME). Also the structure follows the authors’ new paper, Identity Mapping in Deep Residual Networks. — *(b) proposed* in the figure below.

### Like this:

Like Loading...

*Related*

+ It is updated, following their new paper.

LikeLike

Thank you for your work and I’m very interested in implementing this model. Before I start, are there any tips in training this model? Do you do it from scratch, or start with a smaller model? What kind of precision have you attained with this model? Assuming you have a GPU, what kind of training times do you see (or expect)?

LikeLike

Hi, so far I only have the bench mark on MNIST – which is in the readme.md of the repo. Perhaps I could extend it to CIFAR as it’s quite handy to get it in Keras, but not sure I could provide a model that is trained on huge dataset such as ImageNet. Also with my GPU (Tesla with 12GB memory), the author’s model (resnet-151) would not fit.

If there’s a PR that prepare all the stuff I’m willing to run it with my GPU and share more information/trained weight. I would have some vacancy of my GPU soon.

LikeLike

Hello, why do you do the switch of relu and bn before weights

LikeLike

See the figure there, the structure is based on the authors’ update.

LikeLike

Thank you

LikeLike

Is your implementation of resnet50?

LikeLike

No, overall, it is just a arbitrary structure using residual block.

LikeLike