http://caffecn.cn/?/question/509
使用end2end的方法訓練py-faster-rcnn， 把 TRAIN.IMS_PER_BATCH 設置為 2的時候會出錯，顯示data和label的batch size不一致。如下：
在源碼lib/rpn/anchor_target_layer.py中可以看到，anchor_target_layer的top[0] 的batch size被寫死為1了，
這應該就是為什么會出現data 和 label 的batch size不一致錯誤的原因吧吧吧？
但是，為什么anchor_target_layer.py里面的batch size只能是0呢？ 好像官方給的experiments\cfgs\
faster_rcnn_end2end.yml 里面也是把 TRAIN.IMS_PER_BATCH設置為1 了。 是不是end2end訓練的時候，TRAIN.IMS_PER_BATCH只能為1？為什么呢？？
可以怎樣修改代碼使得TRAIN.IMS_PER_BATCH 為任意值嗎？
======================================================================================================================================
Training with Mini-Batch size greater than 1 #267 https://github.com/rbgirshick/py-faster-rcnn/issues/267
how to change the batchsize when training the rpn model?#51 https://github.com/rbgirshick/py-faster-rcnn/issues/51
@rbgirshick:::::It's just not implemented. A reasonable workaround, which is already used, is to set iter_size: N in the solver, in which N is the batch size you want (currently 2 for end to end training). This may be slightly less efficient than batching, but since the inputs are spatially large operating on one image at a time isn't too bad (plus if you use more than one image there's some wasted computation where padding is introduced to fill the 4D tensor).
Batch size of faster rcnn #487 https://github.com/rbgirshick/py-faster-rcnn/issues/487