Patching Weak Convolutional Neural Network Models through Modularization and Composition