Semantic Role Labeling demo 中把lstmemory 换成grumemory后model reload出现准确度问题
Created by: hohaxu
这里只训练2个pass，训练时test accuracy 是0.552631. 但是model reload进来以后test accuracy是0.552867. 具体log信息如下：
==training log== ......I1202 14:09:10.074734 20597 TrainerInternal.cpp:182] Pass=1 Batch=36 samples=5267 AvgCost=105.924 Eval: classification_error_evaluator=0.880542 I1202 14:09:10.419319 20597 Tester.cpp:127] Test samples=5267 cost=102.758 Eval: classification_error_evaluator=0.552631 I1202 14:09:10.419360 20597 GradientMachine.cpp:116] Saving parameters to ./output/pass-00001 I1202 14:09:10.442358 20597 Util.cpp:230] copy ./db_lstm1.py to ./output/pass-00001
==test log== [INFO 2016-12-02 14:09:13,940 networks.py:1282] The input order is [word_data, verb_data, ctx_n1_data, ctx_0_data, ctx_p1_data, mark_data, target] [INFO 2016-12-02 14:09:13,940 networks.py:1289] The output order is [cost_0] I1202 14:09:13.945139 20686 Trainer.cpp:149] trainer: in testing mode I1202 14:09:13.945157 20686 Trainer.cpp:156] trainer mode: Testing I1202 14:09:14.009706 20686 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process I1202 14:09:14.010260 20686 GradientMachine.cpp:127] Loading parameters from ./output/pass-00001 I1202 14:09:15.250123 20686 Tester.cpp:257] Pass=0 samples=5267 AvgCost=102.774 Eval: classification_error_evaluator=0.552867
试了下其他的RNN model，目前看来只有grumemory会有这个问题。 lstmemory，simple_lstm， simple_gru看起来没有这个问题。这是known issue吗？
training script: https://github.com/hohaxu/Paddle/blob/develop/demo/semantic_role_labeling/train_1.sh test script: https://github.com/hohaxu/Paddle/blob/develop/demo/semantic_role_labeling/test_1.sh