# Ch1 build first neutral network
###### tags: `Python`` tensorflow`

## Sample Code

from tensorflow.keras.datasets import mnist

(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

print("train_image[0] : ")
print(train_images[0])

print("\n")

print("train_labels[0] : ")
print(train_labels[0])

print("shape of train_images") # train_image is an array, sized 60000, shaped 28*28
print(train_images.shape)

print("shape of train_labels") # train_label is an array, sized 60000, train_label[0] is 5
print(train_labels.shape)

print("\n")
print("\n")



import matplotlib.pyplot as plt

plt.gcf().set_size_inches(15, 4) # get the current figure
for i in range(5):
    ax = plt.subplot(1, 5, 1+i) # set subplot sized 1 * 5
    ax.imshow(train_images[i], cmap= 'gray') # set output picture gray scale(灰階)
    ax.set_title('label = '+str(train_labels[i]), fontsize=18) # set title
plt.show() # show picture



x_train = train_images.reshape((60000, 28 * 28)) # transform shape of data from (60000, 28, 28) to (60000, 784)
x_train = x_train.astype('float32') / 255 # transform value of pixel from 0-255 to float number between 0-1

x_test = test_images.reshape((10000, 28 * 28))
x_test = x_test.astype('float32') / 255

from tensorflow.keras.utils import to_categorical

y_train = to_categorical(train_labels)
y_test = to_categorical(test_labels)

print("before One-hot coding, label: ")
print(train_labels[0])
print("after One-hot coding, label: ")
print(y_train[0])

print("\n")

print('shape of train_labels :')
print(train_labels.shape)
print("\n")
print('shape of test_labels :')
print(test_labels.shape)

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense

model = Sequential()
model.add(Dense(512, activation='relu', input_dim= 784)) # relu layer of CNN
model.add(Dense(10, activation='softmax'))               # fully connected layers of CNN

model.compile(optimizer='rmsprop',             # optimizer
			  loss='categorical_crossentropy', # loss function
			  metrics=['acc'])                 # benchmark standard ('acc' is abbreviation of 'accuracy')

model.summary()

history = model.fit(x_train, y_train, epochs=5, batch_size=128) # x_train => training sample, y_train => training label(correct answer), epoch => training 5 times, batch_size => how many samples to train in one time

test_loss, test_acc = model.evaluate(x_test, y_test)
print('accuracy to testing dataset : ', test_acc)

predict = model.predict(x_test) # predict() will return predicting result
predict.round(1) # decimal point 1

predict = model.predict_classes(x_test)
print(predict)
print(test_labels)

predict = model.predict_classes(x_test)

plt.gcf().set_size_inches(15, 4)
for i in range(5):
	ax = plt.subplot(1, 5, 1+i)
	ax.imshow(test_images[i], cmap='binary')
	ax.set_title('label = '+str(test_labels[i]) +
				 '\npredi = '+str(predict[i]), fontsize=18)
	ax.set_xticks([]); ax.set_yticks([])
plt.show()

model.save('MnistModel.h5') #save model to file

## Description of Sample Code

* line.1: mnist(Mixed National Institute of Standards and Technology database => we download handwritten dataset from mnist database

* line.24: matplotlib is a plotting(繪圖) library for python programming and NumPy of python

* line.43 & line.45: why do we need to transform value of pixel from 0~255 to 0~1(float), check relu in line.79 and softmax in line.80

* line.50~line.53: take number as example (0-9), train_label[0] is 5; after One-hot coding, result will become[0., 0., 0., 0., 0., 1., 0., 0., 0., 0.]

* line.59: Sequential groups a linear stack of layers into a tf.keras.Model, and it provides training and inference features on this model **(???)**

* line.60: class Dense => densely-connected NN(neutral-network) layer.

* line.62~line.64:
* input layers => let you input values of features (a little bit like linear regression)
* hidden layers => used to do calculating, and they're made by neutrons, and they get data by 'forward propagation'
* output layers => can get result by input_data and calculating by hidden layers

* line.63:
* 512 means number of neutral networks
* relu means Rectified Linear Unit, a kind of activation function in AI
* 784 means amount of dimension, and we only have to type dimensions one time, the following layer would follow dimensions declared in first layer

* line.64:
* 10 means number of neutral networks, it also means output dimensions(0~9)
* softmax is a kind of activation function in AI

* line.64(about softmax):
* three key point about softmax =>
* 1. output value muse between 0 and 1
* 2. total of output value equals to 1
* 3. negative number(-) will become to positive number(+) by itself

take arabic numerals as example
there are ten options: 0 1 2 3 4 5 6 7 8 9
there's a picture
with arabic numerals
result of softmax may become 0.0 0.0 12.0 0.0 0.0 88.0 0.0 0.0 0.0 0.0
so the result of that picture may be 5

* line.70(about model.summary()):
## Sample Code

from tensorflow.keras.datasets import mnist
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.models import load_model # load load_model library

model = load_model('MnistModel.h5') # load model from file
(_,_), (test_images, test_labels) = mnist.load_data()
x_test = test_images.reshape((10000, 28 * 28))
x_test = x_test.astype('float32') / 255
y_test = to_categorical(test_labels)

test_loss, test_acc = model.evaluate(x_test, y_test)

print('accuracy of test data : ', test_acc)

## source
1. https://kknews.cc/zh-tw/code/3342l3o.html (mnist)
2. https://waternotetw.blogspot.com/2018/03/keras-mnist.html (mnist)
3. https://zh.wikipedia.org/wiki/Matplotlib (matplotlib)
4. https://www.itread01.com/content/1545049123.html (to_categorical) (one-hot coding)
5. https://keras.io/api/models/sequential/ (Sequential)
6. https://keras.io/api/layers/core_layers/dense/ (Dense)
7. https://vimsky.com/zh-tw/examples/detail/python-method-keras.layers.Dense.html (Dense)
8. https://zh.wikipedia.org/wiki/%E7%BA%BF%E6%80%A7%E6%95%B4%E6%B5%81%E5%87%BD%E6%95%B0 (ReLU)
9. https://zh.wikipedia.org/wiki/Softmax%E5%87%BD%E6%95%B0 (softmax)
10. https://blog.csdn.net/bitcarmanlee/article/details/82320853 (softmax)

> Written with [StackEdit](https://stackedit.io/).