python - CNTK运行时错误

Question

我正在 cntk 中尝试一个简单的 lstm 网络，但出现以下错误:

RuntimeError                              Traceback (most recent call last)
<ipython-input-58-d0a0e4f580aa> in <module>()
      6         trainer.train_minibatch({x: x1, l: y1})
      7     if epoch % (EPOCHS / 10) == 0:
----> 8         training_loss = trainer.previous_minibatch_loss_average
      9         loss_summary.append(training_loss)
     10         print("epoch: {}, loss: {:.5f}".format(epoch, training_loss))

C:\Program Files\Anaconda3\envs\python2\lib\site-packages\cntk\train\trainer.pyc in previous_minibatch_loss_average(self)
    285         The average training loss per sample for the last minibatch trained
    286         '''
--> 287         return super(Trainer, self).previous_minibatch_loss_average()
    288 
    289     @property

C:\Program Files\Anaconda3\envs\python2\lib\site-packages\cntk\cntk_py.pyc in previous_minibatch_loss_average(self)
   2516 
   2517     def previous_minibatch_loss_average(self):
-> 2518         return _cntk_py.Trainer_previous_minibatch_loss_average(self)
   2519 
   2520     def previous_minibatch_evaluation_average(self):

RuntimeError: There was no preceeding call to TrainMinibatch or the minibatch was empty.

[CALL STACK]
    > CNTK::Trainer::  PreviousMinibatchLossAverage
    - 00007FFFA932A5F6 (SymFromAddr() error: Attempt to access invalid address.)
    - PyCFunction_Call
    - PyEval_GetGlobals
    - PyEval_EvalFrameEx
    - PyEval_GetFuncDesc
    - PyEval_GetGlobals
    - PyEval_EvalFrameEx
    - PyEval_EvalCodeEx
    - PyFunction_SetClosure
    - PyObject_Call (x2)
    - PyObject_CallFunction
    - PyObject_GenericGetAttrWithDict
    - PyType_Lookup
    - PyEval_EvalFrameEx

相关代码为:

# train
loss_summary = []
start = time.time()
for epoch in range(0, EPOCHS):
    for x1, y1 in next_batch(x_train, y_train):
        trainer.train_minibatch({x: x1, l: y1})
    if epoch % (EPOCHS / 10) == 0:
        training_loss = trainer.previous_minibatch_loss_average
        loss_summary.append(training_loss)
        print("epoch: {}, loss: {:.5f}".format(epoch, training_loss))

现在，我已经被困在这个问题上几个小时了，无法理解发生了什么。我正在关注 https://notebooks.azure.com/cntk/libraries/tutorials/html/CNTK_106A_LSTM_Timeseries_with_Simulated_Data.ipynb 上的教程并且搜索谷歌也没有帮助。

感谢您的帮助。

Answer 1

只是一个想法:您的 for(下一个小批量)循环是否可能永远不会执行？

我会尝试使用 pdb 来调试它。只需在 jupyter 单元顶部 import pdb 并在 for x1, y1 .. 循环之前添加 pdb.set_trace() 即可。运行单元格。您可以使用步骤 (s) 进入方法或使用下一步 (n) 继续。这也许可以帮助您分析跟踪，并且您可以使用 pdb 中的打印来证明变量。