接下來就是常見的圖像辨識，TensorFlow有提供一組小圖示包含10種衣服配件類型的圖庫，如下圖所示以及下表所列，包含6000筆的訓練資料以及1000筆的測試資料，每張圖是28x28像素，256灰階。

10種類型分別為：
索引 名稱
0 T-shirt/top
1 Trouser
2 Pullover
3 Dress
4 Coat
5 Sandal
6 Shirt
7 Sneaker
8 Bag
9 Ankle boot

我們先將圖庫load進來，並用print和matplotlib看是長怎樣：

import tensorflow as tf

# Load the Fashion MNIST dataset
fmnist = tf.keras.datasets.fashion_mnist

# Load the training and test split of the Fashion MNIST dataset
(training_images, training_labels), (test_images, test_labels) = fmnist.load_data()

import numpy as np
import matplotlib.pyplot as plt

# You can put between 0 to 59999 here
index = 0

# Set number of characters per row when printing
np.set_printoptions(linewidth=320)

# Print the label and image
print(f'LABEL: {training_labels[index]}')
print(f'\nIMAGE PIXEL ARRAY:\n {training_images[index]}')

# Visualize the image
plt.imshow(training_images[index], cmap='gray_r', vmin=0, vmax=255)

可以看到其中training data中，index為0的圖，而它的類型為9，是"Ankle boot"：

然後就是一樣設定模型和跑training，先將資料normalize除256階，在model部分，前後的layer分別對應到輸入的28x28筆資料展開和輸出的10種類型，中間多了所謂的hidden layer，另外由於有測試資料，所以可用model.evaluate跑看準確度：

# Normalize the pixel values of the train and test images
training_images  = training_images / 255.0
test_images = test_images / 255.0

hidden_layer_num = 128
epochs_num = 5

# Build the classification model
model = tf.keras.models.Sequential([tf.keras.layers.Flatten(), 
                                    tf.keras.layers.Dense(hidden_layer_num, activation=tf.nn.relu), 
                                    tf.keras.layers.Dense(10, activation=tf.nn.softmax)])

#Sequential: That defines a sequence of layers in the neural network.
#Flatten: Flatten just takes that square and turns it into a 1-dimensional array.
#Dense: Adds a layer of neurons
#ReLU effectively means: if x > 0: return x; else: return 0;
#Softmax takes a list of values and scales these so the sum of all elements will be equal to 1.

model.compile(optimizer = tf.optimizers.Adam(),
              loss = 'sparse_categorical_crossentropy',
              metrics=['accuracy'])

fit_history = model.fit(training_images, training_labels, epochs=epochs_num)
print(fit_history.history)

# Evaluate the model on unseen data
eva_history = model.evaluate(test_images, test_labels)
print(eva_history)

我做了張epochs次數在不同hidden layer層數時，對應的訓練模型的loss以及測試的accuracy圖，結果趨勢雖然如同預期，次數或層數越多，loss越低，accuracy越高，但epochs次數越多，或是hidden layer層數越高，訓練的時間也越長，在這例子layer層數為512和1024的loss和accuracy已經沒有顯著地分別了。

另外我們可以掛上一個callback，讓它在訓練到loss低於目標值就停，然後在原本的model.fit加上這個callback：

class myCallback(tf.keras.callbacks.Callback):
  def __init__(self, loss_threshold):
      self.loss_threshold = loss_threshold
      
  def on_epoch_end(self, epoch, logs={}):
    # Check accuracy
    if(logs.get('loss') < self.loss_threshold):

      # Stop if threshold is met
      print("\nLoss is lower than", self.loss_threshold , "so cancelling training!")
      self.model.stop_training = True

# Instantiate class
callbacks = myCallback(0.2)

# Train the model with a callback
model.fit(x_train, y_train, epochs=20, callbacks=[callbacks])

跑到第13次loss已經小於設定的0.2，就自動停了:

Epoch 13/20
1867/1875 [============================>.] - ETA: 0s - loss: 0.1965 - accuracy: 0.9257
Loss is lower than 0.2 so cancelling training!

最後我們可以用predict的方式，看單張圖怎麼運作。結果如下，我們會得到一個10筆數值的list，其中9筆都接近於0，1筆接近於1，這筆的索引就對應到它的類型編號：

classifications = model.predict(test_images)
print(classifications[0])

[4.3171408e-08 8.3716550e-08 8.7536189e-10 8.0485379e-10 1.6189805e-08 2.5948887e-03 4.0446093e-08 8.8888472e-03 1.3672738e-08 9.8851609e-01]

print(test_labels[0])

9