Python learning notes (4)

Xavier:

1. The basic idea is that through the network layer , The variance of input and output is the same , Including forward propagation and backward propagation .

2. If the initialization value is very small , So with the transfer of layers , Variance tends to 0, Enter the value It's getting smaller and smaller , stay sigmoid It's in the 0 near , Close to linear , Without nonlinearity

3. If the initial value is large , So with the transfer of layers , The variance will increase rapidly , At this point, the input value becomes large , and sigmoid Writing a reciprocal at a large input tends to 0, Back propagation will encounter the problem of gradient disappearance

4. Sense and BN The purpose of using is similar

1. np.newaxis： Insert new dimension

2. list and ndarray Interturn ：

   1. list turn numpy：np.array(a)

   2. ndarray turn list：a.tolist()

3. find list Middle front k A relatively large number ：heapq.nlargest(k,list)

4. numpy.ndarray And string to string ：

   1. str=arr.tostring()

   2. arr=np.frombuffer(string, dtype=np.float32)

1. pydub Learning reference for ：https://blog.csdn.net/Debatrix/article/details/59058762

2.  pyworld The use of reference ：https://blog.csdn.net/m0_43395719/article/details/107930075