Write it at the front

For human justice and glory , I've been on a whim recently , I want to update my learning experience of deep learning every day in this blog , One hundred days .
Its utilitarian purpose is to urge oneself to learn , Keep learning ,so,let’s start it!

【1】-2022 year 7 month 15 Japan - Construct a CNN

import keras
from keras.datasets import mnist #  Data sets 
from keras.models import Sequential #  Used to instantiate a class 
from keras.layers import Activation,Dense,Flatten,Conv2D,MaxPooling2D,Dropout # keras Each module of 
from sklearn.utils import shuffle #  Scramble the data 
from sklearn.preprocessing import StandardScaler #  normalization 
from keras.layers import SpatialDropout2D # Drop Upgraded version of layer 
import time,os  #  Used for time statistics and system import 
from keras_flops import get_flops #  Floating point number calculation module 

The structure of the first convolutional neural network is very simple

n_hidden_1 = 64 #  Set hidden layer 
n_classes = 1 #  Set the last output layer 
training_epochs = 5 #  The number of iterations 
batch_size = 100 #  Set the number of batch pictures each time 


model = Sequential()
model.add(Dense(n_hidden_1,activation='relu',input_shape=(42,)))
model.add(Dense(32,activation='relu'))
model.add(Dense(n_classes,activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam',
             metrics=[ 'accuracy'])

print(model.summary()) #  Check the network structure parameters and other information 
flops = get_flops(model, batch_size=batch_size) #  Calculate the floating point number of the network （ Computational complexity ）
#  Model training  
history = model.fit(X_train,y_train,batch_size=batch_size,
                    epochs=training_epochs,validation_data=(X_test,y_test))
model.evaluate(X_test,y_test)

【2】-2022 year 7 month 16 Japan - Some parameters explain

About parameter interpretation ：

EPOCHS = 200 Define the time of training speed
BATCH_SIZE = 128 Represents the sample size of an input neural network
VERBOSE = 1 The progress bar shows
N_HIDDEN = 128 Number of neurons
NB_CLASSES = 10 Input
VALIDATION_SPLIT = 0.2 Check or prove the effectiveness of Training II. The amount of data retained （ Segmentation ratio ）

verbose： The log shows
verbose = 0 Output log information for non-standard output stream
verbose = 1 Record for the output progress bar
verbose = 2 For each epoch Output line record

About binary_crossentropy【 Binary cross entropy 】 and categorical_crossentropy【 Classification cross entropy 】

binary crossentropy:

 It is often used in binary classification problems , It is usually necessary to add sigmoid Use it together 

categorical crossentropy:

 It is suitable for multi classification problems , And use softmax As the activation function of the output layer 

【 On the choice of loss function ：binary_crossentropy、categorical_crossentropy、sparse_categorical_crossentropy】

Learn from the blog ： https://blog.csdn.net/qq_35599937/article/details/105608354

 Dichotomous problem ：

     If it's a dichotomous problem , That is, the final result can only be one of the two classifications , Loss function loss Use binary_crossentropy

 Multiple classification problem ：

    For multi classification problems , In choosing the loss function loss when , It mainly depends on how the data is encoded ：

      1. If it is a classification code （one-hot code ）, Then use categorical_crossentropy

         I am right. one-hot The understanding of coding is ：one-hot Coding is when the label is vectorized , Each label is a N Dimension vector （N It's up to you to decide ）, Only one value of this vector is 1, The rest are 0. That is to index integers i Convert to length of N The binary vector of , This vector has only the second i The element is 1, The rest are 0

        Keras There is a built-in method to vectorize labels ：

from keras.utils.np_utils import to_categorical
 
one_hot_train_labels = to_categorical(train_labels)
one_hot_test_labels = to_categorical(test_labels)
      2. If it's integer encoding , Then use sparse_categorical_crossentropy

        My understanding of integer coding is ： Integer encoding is to put all labels into a vector , Each label corresponds to a value of the vector 
————————————————
 Copyright notice ： This paper is about CSDN Blogger 「 Fu Huatao Fu」 The original article of , follow CC 4.0 BY-SA Copyright agreement , For reprint, please attach the original source link and this statement .
 Link to the original text ：https://blog.csdn.net/fu_jian_ping/article/details/107707780

【3】-2022 year 7 month 17 Japan - Make it clear. Keras Two models in ：Sequential and Model

Original address ：https://blog.csdn.net/weixin_39916966/article/details/88049179

• Sequential It is easier to define the network structure , But simple at the same time , It is also disadvantageous to have built a more complex network
• Model Way to build more complex networks , It's a little harder
  Code display ：

Sequential

from keras.models import Sequential
from keras.layers import Dense, Activation

model = Sequential()
model.add(Dense(32, input_shape = (784,)))
model.add(Activation('relu'))
model.add(Dense(10))
model.add(Activation('softmax'))

Model

from keras.layers import Input, Dense
from keras.models import Model

#  Define input layer , Determine input dimensions 
input = input(shape = (784, ))
# 2 Hidden layers , Every one has 64 Neurons , Use relu Activation function , And the upper layer is taken as the parameter 
x = Dense(64, activation='relu')(input)
x = Dense(64, activation='relu')(x)
#  Output layer 
y = Dense(10, activation='softmax')(x)
#  Defining models , Specify input and output 
model = Model(input=input, output=y)
#  Compile model , Specify optimizer , Loss function , Measure 
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
#  Model fitting , Training 
model.fit(data, labels)

【4】-2022 year 7 month 18 Japan - Define a simple neural network [ be based on TensorFlow2.0]

【5】-2022 year 7 month 19 Japan - Yesterday's accurate baseline improvement strategy of the network 1

【6】-2022 year 7 month 20 Japan - Use random deactivation to further improve

【7】-2022 year 7 month 21 Japan -

【8】-2022 year 7 month 22 Japan -

【9】-2022 year 7 month 23 Japan -