This is an unofficial implementation of clcNet in Keras and Tensorflow. The goal of this project is to replicate the following paper on clcNet architecture.
Following the implementation and hyper-parameter search, benchmarking will be done with respect to
- MobileNet
- ShuffleNet
- SqueezeNet
Disclaimer: The views expressed and/or work on this project done are of my own and do not necessarily represents any of my employeer, organisation, publishing committe or individual related to this research.
A channel local convolution represents a local receptive field for a several channels in the input. Locality of convolution is ensured by application of convolution to a part of input channel to clc block.
The paper proposes a structured way of composing group convolution such that full channel receptive field in a given block is 100 percent. This introduces the idea of more generalized channel local convolution and acyclic graph called channel dependency graph that connects output channel to input channel in a convolution block and measure channel receptive field of that convolution block.
- Observation such as having full channel receptive field in a convolution block is required to have efficient information flow for better representational power in a CNN.
- Interlaced group convolution, one of the two building block for cnc net is developed which with group convolution completes a convolution block.
- Determination of block parameters such as number of groups in IGC and GC is done through minimizing a developed cost function per convolution block remaining under the full channel receptive field constraints.
TO BE UPDATED
TO BE UPDATED
Contribution for this project is currently not acceptable. Notification for any open contribution will be updated here.
Official implementation for clcNet in torch can be found here: