4、 Application of one hot and loss function

---

Preface

I have nothing to do today , Because there are many problems in learning AI , Entanglement starts from that problem ？ But don't worry , I already have the answer . therefore , I decided to summarize the main points of recent learning , To ease the tense learning atmosphere . Ha ha ha ~~, Come on , Get to the point .

/font>

---

Tips ： Let's talk about data preprocessing One-Hot. The following is the main body of this article , The following cases can be used for reference

One 、One-hot What is it? ？

Example ： For this question , Google before , It also involves registers （one-hot Encoding is N The bit status register is N A way to encode States ）.. Really speechless . I won't talk about those at the bottom , We just need to understand one-hot Coding is a form of transforming category variables into machine learning algorithms that are easy to deal with ！
The concept is too abstract , Yes, too much , Let's illustrate with practical examples .
We have two characteristics as follows ：

We see two special evidences named ：animal And food, Explain the meaning of two columns of values , The first column represents the names of animals , The second column is the number of food , Like the first line cat 2 Describe the cat eating two foods , Here is the test data , Mainly want to pass , These data give intuitive understanding and practical operation .
And do the above data one-hot The result after coding is ：

animal The column data type is string , And the second column is numerical , If we can use these eigenvalues 0/1 Express , Is it in machine learning , It is very helpful for these discontinuous values . Sum up , We infer , If you are processing data , Through the step of Feature Engineering , Be able to distinguish the types of features , What is continuous , Which are discontinuous , Then we can deal with it in a special way , Like here one-hot code ！

Two 、 Application scenarios

In model training , The training effect is often regarded as good or bad according to the size of the loss function . You usually use :

1. Mean absolute error (MAEloss),
2. Mean square error (MSEloss), Need to do one-hot And joining softmax Output function .
3. Two categories cross entropy (BCELoss), Need to do one-hot And joining softmax Output function .
4. Two categories cross entropy （BCEWithLogitsLoss, Do it automatically for the input value sigmoid, But it needs to be done one-hot.
5. Multi category cross entropy (CrossEntropyLoss), Automatic processing of input data one-hot and softmax.

3、 ... and 、One-hot Create method (3 Kind of )

1.for Cycle generation

The code is as follows （ Example ）：

 def one_hot(w,h,arr):
     zero_arr = np.zeros([w,h])#w That's ok ,h Column 
     for i in range(w):
         j = int(arr[i])
         zero_arr[i][j]=1
     return zero_arr
 arr = np.array([5, 2, 8, 6])
 one_hot(len(arr),max(arr)+1,arr)

2.arange Traversal methods

The premise of this method is ：arr = np.array([5, 2, 8, 6]) This is the type , Array type .
The code is as follows （ Example ）：

# Law 1
arr = np.array([5, 2, 8, 6])
zero_arr = torch.zeros(len(arr),max(arr)+1)
zero_arr[torch.arange(len(arr)),arr]=1   #arange Very similar range function 
# Or in one step 
# Law 2
torch.zeros(len(arr), max(arr) + 1)[torch.arange(len(arr)),arr]=1

# First level title

3.scatter_ Method

The premise of this method is ：arr = torch.tensor([5, 2, 8, 6]) This is the type , That is, tensor (tensor) type .
The code is as follows （ Example ）：

arr = torch.tensor([5, 2, 8, 6])
    torch_out =torch.zeros(len(arr),max(arr)+1).scatter_(1,arr.reshape(-1,1),1)
'''scatter_(input, dim, index, src) It can be understood as placing elements or modifying elements 
dim： Along which dimension is the index .  It's usually 1 or -1
index： be used for  scatter  The element index of 
src： What content is used to modify the exchange , It can be a scalar or a tensor '''
'''reshape(-1,1) convert to 1 Column ：
reshape(1,-1) Turn it into 1 That's ok ：for
reshape(2,-1) Convert to two lines ：
reshape(-1,2) Convert to two columns ''';

Four 、 summary

Tips ： Here is a summary of the article ：
1. Multi category cross entropy CrossEntropyLoss(), Don't need to do ong-hot, Do it automatically onehot and softmax.
2. Mean square error MSELoss, Need to do onehot And output functions .
3. Two categories cross entropy BCELoss(), Need to do onehot And output functions .
4. Two categories cross entropy BCEWithLogitsLoss,onehot, But there is no need to do output function .