1，報錯Check failed: label < num_classes (5 vs. 5)
打開train_tv_logo.prototxt搜索num_classes改爲9（8classes+1）
2，報錯Check failed: num_priors_ * num_classes_ == bottom[1]->channels() (62136 vs. 34520) .
理解：在multibox_loss_layer裏，有Number of priors must match number of confidence predictions這一句話，查到http://blog.csdn.net/u012235274/article/details/52212077裏面：
預測loction bottom[0] dimension is [N*C*1*1],confidence bottom[1] dimension is [N*C*1*1]
//priors bottom[2] dimension is [N*1*2*W], gound truth bottom[3] dimension is [N*1*H*8]
修改：train_tv_logo.prototxt1148行num_output=36和另外三個，36=4*9,4是一個cell裏面的priorbox數目，9是8class+背景class
3，報錯：Cannot copy param 0 weights from layer ‘fire8_norm_mbox_conf’; shape mismatch. Source param shape is 20 512 3 3 (92160); target param shape is 36 512 3 3 (165888). To learn this layer’s parameters from scratch rather than copying from a saved net, rename the layer.
在train.sh裏最後一行去掉-weights $WEIGHT
4,正常訓練，loss一直保持在6.9左右不下降
初始化有問題
5，test-train時候警告：Missing true_pos for label:
6,改變flip=true後：Check failed: num_priors_ * loc_classes_ * 4 == bottom[0]->channels() (41424 vs. 27616) Number of priors must match number of location predictions
需要把4個mbox_loc層的numoutput從16改成24（4*4->6*4）（test和train都要改）
7，Check failed: num_priors_ * num_classes_ == bottom[1]->channels() (51780 vs. 34520) Number of priors must match number of confidence predictions.
需要把4個mbox_conf層的numoutput從20改成30（4*5->6*5）（test和train都要改）
8，Check failed: 0 == bottom[0]->count() % explicit_count (0 vs. 2) bottom count (80430) must be divisible by the product of the specified dimensions (4)
test裏面reshape裏面的4改成5（test要改，train沒有）