Created by: colin1988

使用gpu训练一个LSTM-CRF模型，CRF解码层无论标识为CPU设备与否都会导致core

I0713 20:41:26.991703 15845 Util.cpp:130] Calling runInitFunctions
I0713 20:41:26.992059 15845 Util.cpp:143] Call runInitFunctions done.
[INFO 2017-07-13 20:41:27,173 networks.py:1466] The input order is [sent_char, sent_seg_flag, target]
[INFO 2017-07-13 20:41:27,173 networks.py:1472] The output order is [crf]
I0713 20:41:27.178159 15845 Trainer.cpp:170] trainer mode: Normal
*** Aborted at 1499949687 (unix time) try "date -d @1499949687" if you are using GNU date ***
PC: @           0x6bb953 paddle::CRFDecodingLayer::init()
*** SIGSEGV (@0x20) received by PID 15845 (TID 0x7f55e9529840) from PID 32; stack trace: ***
    @     0x7f55e9104160 (unknown)
    @           0x6bb953 paddle::CRFDecodingLayer::init()
    @           0x625477 paddle::NeuralNetwork::init()
    @           0x61a131 paddle::MultiGradientMachine::MultiGradientMachine()
    @           0x63d75f paddle::GradientMachine::create()
    @           0x73fa5b paddle::TrainerInternal::init()
    @           0x73c510 paddle::Trainer::init()
    @           0x5b8a49 main
    @     0x7f55e7d21bd5 __libc_start_main
    @           0x5cdfa5 (unknown)
/home/aducz/paddle/paddle_internal_release_tools/idl/paddle/output/bin/paddle_local: line 109: 15845 Segmentation fault      (core dumped) ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_trainer ${@:2}

相关模型配置如下：

 if not is_predict:
     crf_l = crf_layer(
         name='crf',
         size=label_dict_len,
         input=feat,
         label=target,
         param_attr=ParameterAttribute(
             name='crfw', learning_rate=mix_hidden_lr))
          
     crf_dec_l = crf_decoding_layer(
         name='crf_dec_l',
         size=label_dict_len,
         input=feat,
         label=target,
         layer_attr=ExtraAttr(device=-1),
         param_attr=ParameterAttribute(name='crfw'))
          
     eval = sum_evaluator(input=crf_dec_l)
          
     outputs(crf_l)

版本信息

PaddlePaddle 0.9.0, compiled with
    with_avx: ON
    with_gpu: ON
    with_double: OFF
    with_python: ON
    with_rdma: OFF
    with_glog: ON
    with_gflags: ON
    with_metric_learning: 
    with_timer: OFF
    with_predict_sdk: