As anyone aware if there exists some kind of a repository (or some single examples) of deep neural networks models that can be reused for transfer learning? I am aware of one: 'http://s3.amazonaws.com/GraphLab-Datasets/deeplearning/imagenet_model_iter45, However it only can be used as far as I'm concerned with GraphLab Create, not with open source scikit-learn.
@Jack - yes, I did. However it is rather a framework, but without the already learnt networks for reuse in transfer learning approach. At least I could not find it on their website.
Torch7 deep learning framework has one:
https://github.com/soumith/net2net.torch
Here is something I've found, for anyone that came across this post. Tensor flow actually published a tutorial on how to do this.
Looks nice: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/g3doc/how_tos/image_retraining/index.md
And here: https://www.tensorflow.org/versions/r0.8/how_tos/image_retraining/index.html
Caffe framework has a pretty rich repository of models called Model Zoo: https://github.com/BVLC/caffe/wiki/Model-Zoo
Plenty of models are available for transfer learning, from the classical ImageNet winners to more niche architectures trained on specific datasets.
There is also one for Mathematica: https://resources.wolframcloud.com/NeuralNetRepository
which can directly be accessed with the NetModel function.
- Badges
- Science topic