Convolutional Neural Networks with large capacities usually have significant redundancy among different filters and feature channels. By removing this redundancy, we can compress CNNs and decrease their inference time without degrading the performance notably. I decided to prune a network in the medical field to reduce its inference costs. However, network pruning is challenging in this field because pruning should not degrade the performance of models. As a result, it is vital to choose unimportant parts of networks correctly. This is why I employed the genetic algorithm to identify redundant filters of my U-Net-based network. I also tested other criteria to determine unimportant parts of my network, such as APoZ and L2norm. I used the pruned model for brain tumor segmentation to check the effectiveness of my technique. To read more details about the brain tumor segmentation project, click here.
- Training my network on multimodal MRIs
- Encoding desired layers of the trained network into chromosomes
- Finding less important filters of each layer using the genetic algorithm
- Removing identified filters
- Retraining the pruned model