Caffe
A deep learning framework designed for expression, speed, and modularity
&
| + | Expressive Architecture | Network design using configuration files. |
|---|---|---|
| + | Modularity | Swap and combine layers for custom architectures. |
| + | Efficiency | Optimized for fast training and inference on CPU/GPU. Leverages cuDNN library for GPU acceleration and Intel Math Kernel Library for CPU acceleration. |
| + | Pre-trained Models | Access to popular models like AlexNet, VGGNet, and GoogLeNet. |
| + | Support for Various Architectures (CNN, RCNN, LSTM) | Supports CNN (Convolutional Neural Network) for deep learning, RCNN (Region-based CNN) used for object detection and localization, LSTM (Long Short-Term Memory) suitable for sequence modeling. |
| - | Compilation Complexity | Requires manual compilation from source code for deployment, which can be troublesome due to hardware dependencies. |
| - | Learning Curve | May not be intuitive for beginners, especially those unfamiliar with C++ or configuration-based model definitions. |
| - | New Layers | New network layers must be coded in C++ or CUDA, limiting flexibility |
| - | Narrow Model Coverage | May not cover newer deep learning architectures not already included in Caffe |
| - | Distributed Computing | Does not support distributed computing. |
| - | Integration Limitations | Integration with other tools and frameworks might be less smooth than with more popular options. |
| - | Lack of Active Development | Limited future support. |
System Requirements
Version ↓
| # | Minimum | Recommended | Optimal |
|---|---|---|---|
| 1 | CUDA v5.5, and 5.0 (considered legacy) | CUDA v6.* | CUDA v7+ |
| 2 | Basic Linear Algebra Subprograms via ATLAS, MKL, or OpenBLAS I Boost >= 1.55 I protobuf, glog, gflags, hdf5 | Minimum plus (optional): OpenCV >= 2.4 including 3.0 I IO libraries: lmdb, leveldb (note: leveldb requires snappy) I cuDNN for GPU acceleration (v6) | |
| 3 | For Python Caffe: Python 2.7 or Python 3.3+, numpy (>= 1.7), boost-provided boost.python I For MATLAB Caffe: MATLAB with the mex compiler. |
Written in
C++, Python, CUDA
Initial Release
10 October 2013
Alternatives
Notes
- Caffe: Convolutional Architecture for Fast Feature Embedding