This is a reference implementation of distributed CNN training using fusing, tiling and grouping. It is built upon the original darknet framework. The technique targets early feature dominated convolutional and pooling layers.
The implementation is added as a feature onto the darknet framework. The core implementation of the technique is under src/ftp.c. To enable the feature, under the Makefile, the FTP option can be set to 1 and built. There is a configuration file cfg/ftp.cfg where the parameters for the distributed training setup can be configured, namely the number of tiling grid, number of initial layers, device IP addresses, and the grouping profiles.
Currently, the implementation targets the early layers and is just a prototype implementation. Future work will explore distributed techniques for the later layers which are generally weight dominated to train on the complete network.
If you have any questions or want to discuss anything about this, author can be contacted at
- Pranav Rama - pranavrama9999@utexas.edu