This is an easy to use Torch (re)-implementation of the Convolutional Bag-of-Features (CBoF) pooling method (as presented in Bag-of-Features Pooling for Deep Convolutional Neural Networks). CBoF is a useful tool that allows for decoupling the size of the representation extracted from a deep CNN from both the size and the number of the extracted feature maps. Therefore, it allows for reducing the size of the used CNNs as well as for improving the scale-invariance of the models. This repo is inherited by [Keras implementation of CBoF]https://github.com/passalis/keras_cbof
To use the CBoF pooling method, simply insert the BoF_Pooling layer between the last convolution and a fully connected layer:
from cbof import BoF_Pooling, initialize_bof_layers
...
self.BoF_pooling = BoF_Pooling()
...
initialize_bof_layers(model, x_train)Remember to initialize the BoF layer (the initialize_bof_layers() function automatically initializes all the BoF layers in a torch model). The number of codewords (that defines the dimensionality of the extracted representation) as well as the spatial level must be defined. Two spatial levels are current supported: 0 (no spatial segmentation) and 1 (spatial segmentation in 4 regions).