Cifar10 transfer learning results

Cifar10 is a good dataset for the beginner. And here is the comparison output of the results based on different implementation methods.

The cifar experiment is done based on the tutorial provided by

the first version is exactly the same one as shown in the tutorial

the gpu version is changed from without padding to padding to padding+deeper network

the resnet18 is based on the resnet 18 with and without pretrain also frozen the conv parameters and unfrozen the parameters of the conv layer. detail is given as below:

File Name pretrain epoch frozen conv result no 10 n/a 71 % yes 10 yes 76 % yes 30 yes 77 % yes 30 no 93 %

code is shared below: