Created by: linrongyi

序列标注模型预测的时候, 如果ouput layer是id类型的, 例如 crf_decoding 或者 maxid

    if use_crf:

      Layer(
          name = "%s_label" % name_,
          type = "crf_decoding",
          size = slot_conf_dict[name_]['dim'],
          inputs = [
            Input("output_%s" % name_, parameter_name="crfw_%s" % name_),
          ]
      )

    else:

      Layer(
        name = "%s_label" % name_,
        type = "maxid",
        inputs = "output_%s" % name_,
      )

forwardTest会出错

raceback (most recent call last):
  File "deep_tagger_new_paddle.py", line 87, in <module>
    outputs = do_tag(text, conf='output_tag')
  File "deep_tagger_new_paddle.py", line 57, in do_tag
    outputs_ = networks[0].forwardTest(inst_)
  File "/home/aladdin/.jumbo/lib/python2.7/site-packages/py_paddle/util.py", line 149, in forwardTest
    for i in xrange(outArgs.getSlotNum())
  File "/home/aladdin/.jumbo/lib/python2.7/site-packages/py_paddle/util.py", line 89, in __arguments_to_numpy__
    if ids is not None:
UnboundLocalError: local variable 'ids' referenced before assignment

如果输出是概率分布, e.g.,

  Layer(
      name = "output_%s" % name_,
      type = "mixed",
      active_type = "softmax",
      size = slot_conf_dict[name_]['dim'],
      bias = False,
      inputs = [
          FullMatrixProjection("decoder_state_final_%s" % name_),
      ]
  )

则不会出错. 这是为什么呢?