Finetune问题

大家好，一直很困惑caffe在做fine-tuning时，有两个问题：

0. 测试方法

0.1 **PLDA** ：通过SDK提取不同CNNs（Caffenet、Googlenet）网络指定层输出作为特征，然后用plda对特征进行训练、分类。统计结果以**单字识别率**为指标。
0.2 **SSL**(structure Sparse Learning): 基于Alexnet训练得到的模型，基于上述模型SSL的模型，基于SSL模型finetune得到的调整模型

How many lives do we live?
How many times do we die?
They say we lose 21grams at the exact moment of our death.
Everyone.
And how much fits into 21grams?
How much is lost? When do we lose 21 grams?
How much goes with them? How much is gained?
Twenty-one grams.
The weight of a stack of five nickels.
The weight of a hummingbird.
A chocolate bar.
How much did 21 grams weigh?