Image Recognition (MNIST dataset) in Keras

# tensorflow random seed set to reproduce later
import tensorflow as tf
#tf.random.set_seed(1234) # poor
tf.random.set_seed(1) # poor

1) importing and reading dataset

(train_ft, train_lb), (test_ft, test_lb) = \
                                        tf.keras.datasets.mnist.load_data()
                                        #################

print(f'train_set:test_set={len(train_ft)}:{len(test_ft)}')

train_set:test_set=60000:10000

# plot input data
import matplotlib.pyplot as plt
fig_container = plt.figure(figsize=(20,20)) # .figure is top container
nrows = 1
ncols = 5
for i in range(ncols):
    # subplot(nrows, ncols, index, **kwargs)
    ax = fig_container.add_subplot(nrows, ncols, i+1) 
    ax.imshow(train_ft[i], cmap='Greys')
    ax.set_title('label=' + str(train_lb[i]))

2) preprocessing data

# input flatten
train_ft[0].shape # (28, 28)
train_ft_flat = train_ft.reshape(-1, 28*28)
test_ft_flat = test_ft.reshape(-1, 28*28)

# categorize
num_classes = len(set(train_lb))
print(num_classes)
# one-hot-encode
train_lb_cat = tf.keras.utils.to_categorical(train_lb, num_classes) 
test_lb_cat = tf.keras.utils.to_categorical(test_lb, num_classes) 
               ##############

3) building neural network model

# build a model
model = tf.keras.models.Sequential()
model.add( tf.keras.layers.Dense(128, activation='relu', \
                                      input_shape=(28*28,)) ) # input
model.add( tf.keras.layers.Dropout(0.2) )
model.add( tf.keras.layers.Dense(64, activation='relu') )
model.add( tf.keras.layers.Dropout(0.2) )
model.add( tf.keras.layers.Dense(num_classes, \
                                    activation='softmax') ) # output

# compile a model with specific optimizer, loss, and monitoring metric
model.compile(optimizer=tf.keras.optimizers.Adam(), \
                       #tf.keras.optimizers.RMSprop(0.01), 
              loss=tf.keras.losses.CategoricalCrossentropy(),
              metrics=[tf.keras.metrics.CategoricalAccuracy()])
#model.compile( optimizer='adam', \
                loss='categorical_crossentropy', \
                metrics=['accuracy'])
model.summary()

Model: "sequential_8"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_20 (Dense)             (None, 128)               100480    
_________________________________________________________________
dropout_14 (Dropout)         (None, 128)               0         
_________________________________________________________________
dense_21 (Dense)             (None, 64)                8256      
_________________________________________________________________
dropout_15 (Dropout)         (None, 64)                0         
_________________________________________________________________
dense_22 (Dense)             (None, 10)                650       
=================================================================
Total params: 109,386
Trainable params: 109,386
Non-trainable params: 0
_________________________________________________________________

4) training neural network

# training fit to **train_set** using specific epochs & batch_size
model.fit(train_ft_flat, train_lb_cat, epochs=3, batch_size=20)

Epoch 1/3
3000/3000 [==============================] - 6s 2ms/step - loss: 2.3589 - categorical_accuracy: 0.6836
Epoch 2/3
3000/3000 [==============================] - 6s 2ms/step - loss: 0.6324 - categorical_accuracy: 0.8352
Epoch 3/3
3000/3000 [==============================] - 6s 2ms/step - loss: 0.4446 - categorical_accuracy: 0.8817

<tensorflow.python.keras.callbacks.History at 0x15bfd2550>

5) evaluation model

# prediction using test set
predict_vec = model.predict(test_ft_flat)

import numpy as np
predictions = [np.argmax(p) for p in predict_vec]

# plot prediction vs. true label
print(f'true label={test_lb[4]}')
print(f'predict label={predictions[4]}')
plt.imshow(test_ft[4], cmap='Greys')

true label=4
predict label=4

<matplotlib.image.AxesImage at 0x15532ee80>

# accuracy
correct_num = 0
for i, p in enumerate(predictions):
    if test_lb[i] == p:
        correct_num += 1
print(f'model accuracy = {correct_num/len(predictions)}')

model accuracy = 0.9067