Questions about network training results #25
Description
When the network is in 300 epoch, it feels that it can no longer improve the accuracy in convergence, but the optimal weight provided is obtained in 1195 batches. What details in the source code I didn't notice? The accuracy rate trained by myself is 46.36%, which is 1.74% worse than the data in the paper.