我目前已经由edx注册了语音识别系统。到目前为止，我已经完成了两个模块，并停留在第三个模块上。我已经从给定的音频信号中提取了特征。我正在尝试使用cntk训练具有这些功能的DNN。以下是我遇到的错误。
Selected GPU[0] GeForce 940MX as the process wide default device.
Training 1321080 parameters in 10 parameter tensors.
-------------------------------------------------------------------
Build info:
Built time: Nov 22 2017 22:00:37
Last modified date: Mon Oct 16 23:15:20 2017
Build type: Release
Build target: GPU
With 1bit-SGD: no
With ASGD: yes
Math lib: mkl
CUDA version: 9.0.10
CUDNN version: 6.0.21
Build Branch: HEAD
Build SHA1: feade5b1bc1fbe6a8fc13b214f66aa89dac0db53
MPI distribution: Microsoft MPI
MPI version: 7.0.12437.6
-------------------------------------------------------------------
Redirecting log to file ../Experiments\am\DNN\log
Reading script file ../Experiments\lists\feat_train.rscp ... 1832 entries
HTKDeserializer: selected '1832' utterances grouped into '1' chunks, average chunk size: 1832.0 utterances, 73280.0 frames (for I/O: 1832.0 utterances, 73280.0 frames)
HTKDeserializer: determined feature kind as '584'-dimensional 'USER' with frame shift 10.0 ms
Total (121) state names in state list '../Experiments\am\labels.ciphones'
MLFDeserializer: '2559' utterances with '1809843' frames
Reading script file ../Experiments\lists\feat_dev.rscp ... 466 entries
HTKDeserializer: selected '466' utterances grouped into '1' chunks, average chunk size: 466.0 utterances, 18640.0 frames (for I/O: 466.0 utterances, 18640.0 frames)
HTKDeserializer: determined feature kind as '1466'-dimensional 'USER' with frame shift 10.0 ms
Total (121) state names in state list '../Experiments\am\labels.ciphones'
MLFDeserializer: '2559' utterances with '1809843' frames
attempt: setkind: inconsistent feature kind for file '../Experiments\lists\..\feat\1272-128104-0001.feat', retrying 2-th time out of 5...
attempt: setkind: inconsistent feature kind for file '../Experiments\lists\..\feat\1272-128104-0001.feat', retrying 3-th time out of 5...
attempt: setkind: inconsistent feature kind for file '../Experiments\lists\..\feat\1272-128104-0001.feat', retrying 4-th time out of 5...
attempt: setkind: inconsistent feature kind for file '../Experiments\lists\..\feat\1272-128104-0001.feat', retrying 5-th time out of 5...
Traceback (most recent call last):
File "M3_Train_AM.py", line 283, in <module>
main()
File "M3_Train_AM.py", line 277, in main
model_type=model_type
File "M3_Train_AM.py", line 197, in train_network
cv_source=cv_source
File "M3_Train_AM.py", line 154, in train_and_test
cv_config=cv_checkpoint_config
File "C:\Users\Shivani\Anaconda3\lib\site-packages\cntk\internal\swig_helper.py", line 69, in
wrapper
result = f(*args, **kwds)
File "C:\Users\Shivani\Anaconda3\lib\site-packages\cntk\train\training_session.py", line 333,
in train
super(TrainingSession, self).train(device)
File "C:\Users\Shivani\Anaconda3\lib\site-packages\cntk\cntk_py.py", line 3361, in train
return _cntk_py.TrainingSession_train(self, computeDevice)
RuntimeError: setkind: inconsistent feature kind for file '../Experiments\lists\..\feat\1272-128104-0001.feat'
[CALL STACK]
> CreateDeserializer
- CreateDeserializer (x6)
- CreateCompositeDataReader (x8)
- Microsoft::MSR::CNTK::Matrix<float>:: __autoclassinit2
但是当尝试使用相同的功能训练BLSTM时，以下是我遇到的错误，
Selected GPU[0] GeForce 940MX as the process wide default device.
Training 8691832 parameters in 3 parameter tensors.
-------------------------------------------------------------------
Build info:
Built time: Nov 22 2017 22:00:37
Last modified date: Mon Oct 16 23:15:20 2017
Build type: Release
Build target: GPU
With 1bit-SGD: no
With ASGD: yes
Math lib: mkl
CUDA version: 9.0.10
CUDNN version: 6.0.21
Build Branch: HEAD
Build SHA1: feade5b1bc1fbe6a8fc13b214f66aa89dac0db53
MPI distribution: Microsoft MPI
MPI version: 7.0.12437.6
-------------------------------------------------------------------
Redirecting log to file ../Experiments\am\BLSTM\log
Reading script file ../Experiments\lists\feat_train.rscp ... 1832 entries
HTKDeserializer: selected '1832' utterances grouped into '1' chunks, average chunk size: 1832.0 utterances, 73280.0 frames (for I/O: 1832.0 utterances, 73280.0 frames)
HTKDeserializer: determined feature kind as '584'-dimensional 'USER' with frame shift 10.0 ms
Traceback (most recent call last):
File "M3_Train_AM.py", line 283, in <module>
main()
File "M3_Train_AM.py", line 277, in main
model_type=model_type
File "M3_Train_AM.py", line 177, in train_network
frame_mode=frame_mode)
File "M3_Train_AM.py", line 53, in create_mb_source
return C.io.MinibatchSource([fd, ld], frame_mode=frame_mode, max_sweeps=max_sweeps)
File "C:\Users\Shivani\Anaconda3\lib\site-packages\cntk\io\__init__.py", line 226, in __init__
source = cntk_py.create_composite_minibatch_source(config)
RuntimeError: HTKDeserializer: model vector size is not multiple of input features
[CALL STACK]
> CreateDeserializer
- CreateDeserializer (x3)
- CreateCompositeDataReader (x3)
- CNTK:: UniversalLearner
- CNTK:: CreateCompositeMinibatchSource
- PyInit__cntk_py
- PyCFunction_FastCallDict
- PyObject_CallFunctionObjArgs
- PyEval_EvalFrameDefault
- Py_CheckFunctionResult
- PyFunction_FastCallDict
- PyObject_IsInstance
我也参加了该课程，Jim Lewis提出了一个解决方案。请参阅下面的链接。
看看是否可行！