Elamur Research
☰
home / papers / 2014 / Caffe: Convolutional Architecture for Fast Feature Embedding

Caffe: Convolutional Architecture for Fast Feature Embedding

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv — non-exclusive licence to distribute" does not allow redistributing the text itself. Original — arxiv.org/pdf/1408.5093

Summary

Deep convolutional networks had become the leading approach to visual recognition, but reproducing published results could take months. Existing toolkits rarely combined fast execution with ready-made state-of-the-art models. The UC Berkeley team (BVLC) presents Caffe, an open-source deep learning framework written in C++. It uses CUDA for GPU acceleration and offers Python/Numpy and MATLAB bindings. Networks are declared in Protocol Buffer configuration files as arbitrary directed acyclic graphs of layers, which keeps the model definition independent of the executing code. A single call moves the same network between CPU and GPU. Layers exchange data through 4-dimensional "blobs" that handle host–device synchronization and allocate memory lazily. Every module has unit tests. Training relies on SGD with momentum, learning-rate schedules, snapshots and fine-tuning. Caffe ships reference models with recipes to reproduce them and handles over 40 million images per day on one GPU. Its speed, modularity and focus on reproducibility make it a practical shared base for research and industrial deployment.

Key points

  • Throughput is about 2.5 ms per image (over 40M images/day) on a single NVIDIA K40 or Titan GPU. LevelDB plus Protocol Buffers storage reaches 150 MB/s on commodity machines.
  • Built-in layers include convolution, pooling, inner product, ReLU/logistic, local response normalization, element-wise operations, and softmax and hinge losses. The MNIST example is LeNet (lenet_train.prototxt).
  • The code is BSD-licensed on GitHub (BVLC/caffe). The reference models, such as the "AlexNet" ImageNet model and the R-CNN detector, are restricted to academic and non-commercial use.
  • Table 1 compares Caffe with cuda-convnet, Decaf, OverFeat, Theano/Pylearn2 and Torch7. Its distinguishing features are a pure C++ core with a CPU mode and off-the-shelf pretrained models.
  • Reported applications include:
    • R-CNN with Selective Search, top detection results on PASCAL VOC 2007–2012 and the ImageNet 2013 Detection challenge;
    • fine-tuning to all 10,000 ImageNet categories for open-vocabulary retrieval;
    • style recognition on the Flickr Style dataset.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Other

Cited by

312 papers of our library cite it (of the 140673 whose bibliographies we read); 10 of them have a page here, 302 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1408.5093 · arXiv 1408.5093

@misc{jia2014caffe,
  title = {{Caffe: Convolutional Architecture for Fast Feature Embedding}},
  author = {Jia, Yangqing and Shelhamer, Evan and Donahue, Jeff and Karayev, Sergey and Long, Jonathan and Girshick, Ross and Guadarrama, Sergio and Darrell, Trevor},
  year = {2014},
  doi = {10.48550/arxiv.1408.5093},
  eprint = {1408.5093},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1408.5093}
}

References

11 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. R. Collobert, K. Kavukcuoglu, and C. Farabet. Torch7: A MATLAB-like environment for machine learning. In BigLearn, NIPS Workshop, 2011. find
  2. J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang, E. Tzeng, and T. Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In ICML, 2014. find
  3. R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, 2014. find our card
  4. I. Goodfellow, D. Warde-Farley, P. Lamblin, V. Dumoulin, M. Mirza, R. Pascanu, J. Bergstra, F. Bastien, and Y. Bengio. Pylearn2: a machine learning research library. arXiv preprint 1308.4214, 2013. — [arXiv:1308.4214](https://arxiv.org/abs/1308.4214) find
  5. S. Guadarrama, E. Rodner, K. Saenko, N. Zhang, R. Farrell, J. Donahue, and T. Darrell. Open-vocabulary object retrieval. In RSS, 2014. find
  6. S. Karayev, M. Trentacoste, H. Han, A. Agarwala, T. Darrell, A. Hertzmann, and H. Winnemoeller. Recognizing image style. arXiv preprint 1311.3715, 2013. — [arXiv:1311.3715](https://arxiv.org/abs/1311.3715) find
  7. A. Krizhevsky. cuda-convnet. https://code.google.com/p/cuda-convnet/, 2012. ↗
  8. A. Krizhevsky, I. Sutskever, and G. Hinton. ImageNet classification with deep convolutional neural networks. In NIPS, 2012. find
  9. P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun. Overfeat: Integrated recognition, localization and detection using convolutional networks. In ICLR, 2014. find
  10. J. Uijlings, K. van de Sande, T. Gevers, and A. Smeulders. Selective search for object recognition. IJCV, 2013. find
  11. N. Zhang, M. Paluri, M. Ranzato, T. Darrell, and L. Bourdev. Panda: Pose aligned networks for deep attribute modeling. In CVPR, 2014. https://github.com/BVLC/caffe/http://caffe.berkeleyvision.org/ ↗