This folder contains Residual Network (ResNet) [1] models trained on ImageNet by Marcel Simon at the Computer Vision Group Jena (CVGJ) using the Caffe framework. Each subfolder contains a model of different depth. We use the pre-activation variant and mostly followed the FAIR implementation [2], which is available for Torch. The generation script was based on a previous implementation of Oscar Beijbom [3].

No mean subtraction is required for the pre-trained models! We use a batch-normalization layer which basically does the same.

The pre-trained models can be obtained by the download link written in model_download_link.txt.

The network definition files were generated using the Python 3 script resnet_preact.py. You can generate your own network file using the following Python 3 code:

import resnet_preact
proto = resnet_preact.residual_net(50)
open('train.prototxt','w').write(str(proto.to_proto()))

If you want to train a generated model, copy the solver from our pre-trained models and simply execute caffe train --solver train.solver --gpu 0 2> train.log to start the training and write the output to the log file train.log.

To evaluate the final model, execute caffe train --solver test.solver --gpu 0 2> test.log.

Single-crop error rates on the validation set of the ILSVRC 2012--16 classification task.

Please cite the following technical report if our models helped your research:

@article{simon2016cnnmodels,
  Author = {Simon, Marcel and Rodner, Erik and Denzler, Joachim},
  Journal = {arXiv preprint arXiv:1612.01452},
  Title = {ImageNet pre-trained models with batch normalization},
  Year = {2016}
}

The report also contains an overview and analysis of the models shown here.

[1]: He, Kaiming, et al. "Deep residual learning for image recognition." arXiv preprint arXiv:1512.03385 (2015). [2]: https://github.com/facebook/fb.resnet.torch [3]: https://github.com/beijbom/beijbom_vision_lib