總結一下今天的學習過程

（注：我此刻的心情與剛剛三分鐘前的心情是完全不一樣的）

（昨天在想一些錯誤，今天又重拾信心重新配置GPU環境，結果很失敗，不過現在好了，在尋思著今天干了什麼的時候，無意間想到是不是自己方法入口不對啊。結果果然很幸運的被我猜到了，，，哈哈哈，我的心情又好了）

總結最大的體會：有的時候在程式碼不能執行的時候，可是嘗試先看看學習程式碼，

起碼從程式碼的入口呼叫看起、看清、看準，不要整個全網基本不會出現的錯誤讓自己造出來了，更重要的是自己看不懂這個錯誤

起碼從程式碼的入口呼叫看起、看清、看準，不要整個全網基本不會出現的錯誤讓自己造出來了，更重要的是自己看不懂這個錯誤

起碼從程式碼的入口呼叫看起、看清、看準，不要整個全網基本不會出現的錯誤讓自己造出來了，更重要的是自己看不懂這個錯誤

哎 崩潰的一天

總結一下昨天失敗的原因

1，編碼問題

使用TextEncoding.exe檔案將E:\Python\CRDA\cuda8\include中編碼改成Unicode

2，什麼出現巨集定義啥的，解決的辦法忘了，明天或許會幫同學重灌，到時再補充

3，cuda9與cuda8再vs2015中衝突，vs2015中使用的的是cuda9，將其cuda9解除安裝和cuda8都解除安裝，重新安裝cuda8

4，使用vs2015編譯cudaruntime程式，並執行，生成一些所謂的必要檔案，我也是傻懵的，結果少了很多錯誤，最好在64位和32位都debug編譯

5，在編譯cudaruntime程式時，會出現FIB類似的檔案打不開或者無法找到，這些都可以在百度中尋找答案，建議：不要在意有關nvxxxx.dll檔案找不到無法載入，然後自己又去DOS視窗使用regsvr32 C:WindowsSysWOW64\nvcuda.dll等等，這是非常愚蠢的，，

會出現一個已載入但找不到入口點DLLRegisterServer

只要不是nvxxx.dll的都可以忽略

6，其他的一些程式碼問題，百度解決一下即可

其中關於downsample與pool_2d，由於Python的版本不同，Python3中被替換

pooled_out = pool_2d(input=conv_out, ws=self.poolsize, ignore_border=True)這裡是ws不是ds，不然會有一個警告

 UserWarning: DEPRECATION: the 'ds' parameter is not going to exist anymore as it is going to be replaced by the parameter 'ws'.
  pooled_out = pool_2d(input=conv_out, ds=self.poolsize, ignore_border=True)

conv2d這個庫匯入也要注意，有時候

from theano.tensor.nnet import conv2d

conv_out = conv2d(
            input=self.inpt, filters=self.w, filter_shape=self.filter_shape,
            input_shape=self.image_shape)

7，還有一個嚴重的錯誤是關於cudnn版本的

在cudnn-8.0-windows7-x64-v7中有一個XXX_DV4，類似cudnnGetPoolingNdDescriptor_v4的錯誤，說是找不到此檔案，這裡要將cudnn-8.0-windows7-x64-v7替換成cudnn-8.0-windows7-x64-v5.1，警告也解決了，錯誤也解決了

其他的錯誤一時半會想不出來了，反正就是太多太多了

最後說一下今天的最大的失誤愚蠢

起碼從程式碼的入口呼叫看起、看清、看準，不要整個全網基本不會出現的錯誤讓自己造出來了，更重要的是自己看不懂這個錯誤

造成的錯誤：

Trying to run under a GPU.  If this is not desired, then modify network3.py
to set the GPU flag to False.
(<CudaNdarrayType(float32, matrix)>, Elemwise{Cast{int32}}.0)
cost: Elemwise{add,no_inplace}.0
grads: [Elemwise{add,no_inplace}.0, GpuFromHost.0, Elemwise{add,no_inplace}.0, GpuFromHost.0, Elemwise{add,no_inplace}.0, GpuFromHost.0]
updates: [(<CudaNdarrayType(float32, 4D)>, Elemwise{sub,no_inplace}.0), (<CudaNdarrayType(float32, vector)>, Elemwise{sub,no_inplace}.0), (w, Elemwise{sub,no_inplace}.0), (b, Elemwise{sub,no_inplace}.0), (w, Elemwise{sub,no_inplace}.0), (b, Elemwise{sub,no_inplace}.0)]
train_mb: <theano.compile.function_module.Function object at 0x000000002290B0B8>
Training mini-batch number 0
Traceback (most recent call last):
  File "E:\Python\Anaconda3\lib\site-packages\theano\compile\function_module.py", line 884, in __call__
    self.fn() if output_subset is None else\
ValueError: GpuReshape: cannot reshape input of shape (10, 20, 12, 12) to shape (10, 784).

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "E:\Python\NewPythonData\neural-networks-and-deep-learning\src\demo3.py", line 38, in <module>
    net.SGD(training_data,10,mini_batch_size,0.1,validation_data,test_data)
  File "E:\Python\NewPythonData\neural-networks-and-deep-learning\src\network3.py", line 179, in SGD
    train_mb(minibatch_index)
  File "E:\Python\Anaconda3\lib\site-packages\theano\compile\function_module.py", line 898, in __call__
    storage_map=getattr(self.fn, 'storage_map', None))
  File "E:\Python\Anaconda3\lib\site-packages\theano\gof\link.py", line 325, in raise_with_op
    reraise(exc_type, exc_value, exc_trace)
  File "E:\Python\Anaconda3\lib\site-packages\six.py", line 685, in reraise
    raise value.with_traceback(tb)
  File "E:\Python\Anaconda3\lib\site-packages\theano\compile\function_module.py", line 884, in __call__
    self.fn() if output_subset is None else\
ValueError: GpuReshape: cannot reshape input of shape (10, 20, 12, 12) to shape (10, 784).
Apply node that caused the error: GpuReshape{2}(GpuElemwise{add,no_inplace}.0, TensorConstant{[ 10 784]})
Toposort index: 51
Inputs types: [CudaNdarrayType(float32, 4D), TensorType(int32, vector)]
Inputs shapes: [(10, 20, 12, 12), (2,)]
Inputs strides: [(2880, 144, 12, 1), (4,)]
Inputs values: ['not shown', array([ 10, 784])]
Outputs clients: [[GpuElemwise{Composite{(scalar_sigmoid(i0) * i1)},no_inplace}(GpuReshape{2}.0, GpuFromHost.0)]]

HINT: Re-running with most Theano optimization disabled could give you a back-trace of when this node was created. This can be done with by setting the Theano flag 'optimizer=fast_compile'. If that does not work, Theano optimizations can be disabled with 'optimizer=None'.
HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node.

改正後，結果非常激動

Trying to run under a GPU.  If this is not desired, then modify network3.py
to set the GPU flag to False.
(<CudaNdarrayType(float32, matrix)>, Elemwise{Cast{int32}}.0)
cost: Elemwise{add,no_inplace}.0
grads: [Elemwise{add,no_inplace}.0, GpuFromHost.0, Elemwise{add,no_inplace}.0, GpuFromHost.0, Elemwise{add,no_inplace}.0, GpuFromHost.0]
updates: [(<CudaNdarrayType(float32, 4D)>, Elemwise{sub,no_inplace}.0), (<CudaNdarrayType(float32, vector)>, Elemwise{sub,no_inplace}.0), (w, Elemwise{sub,no_inplace}.0), (b, Elemwise{sub,no_inplace}.0), (w, Elemwise{sub,no_inplace}.0), (b, Elemwise{sub,no_inplace}.0)]
train_mb: <theano.compile.function_module.Function object at 0x0000000022A18208>
Training mini-batch number 0
Training mini-batch number 1000
Training mini-batch number 2000
Training mini-batch number 3000
Training mini-batch number 4000
Epoch 0: validation accuracy 93.73%
This is the best validation accuracy to date.
The corresponding test accuracy is 93.20%

scikit-neuralnetwork中的examples/plot_mlp.py

測試技測試成功

python examples/plot_mlp.py --params activation

結果