DNNs are defined in Caffe layer-by-layer. Layer is the essence of a model and the fundamental unit of computation. Data enters Caffe through data layers. Accepted data sources are efficient databases (LevelDB or LMDB), memory, file system, Hierarchical Data Format (HDF5) or common image formats (e.g., GIF, TIFF, JPEG, PNG, PDF). Common and normalization layers provide various data vector processing and normalisation operations. New layers must be written in C++/CUDA, although custom layers are also supported in Python (but are less efficient).

Strong points

• Suitable for FFNN and excellent implementation of CNN for image processing.
• Fastest DL library on CPU, GPU out-of-the-box training.
• A number of pre-trained networks directly from the Caffe Model Zoo, available for immediate use.
• Easy to code (API/CLI) with Python and MatLab interface.
• Well-acceptable from research community.

Weak points

• It is not good RNN i.e. for text, sound and time-series data.
• Cumbersome for complicated DNN models i.e. GoogleLeNet and ResNet.
• Custom layers must be written in C++.

Return to Contemt