3.3 手写体识别

问题描述

识别0-25这26个数，数据是csv格式，即对应数字的像素分布。

数据预处理

import csv
import numpy as np
import tensorflow as tf
from tensorflow.keras.preprocessing.image import ImageDataGenerator

def get_data(filename):
    with open(filename) as training_file:
        csv_reader = csv.reader(training_file, delimiter=',')
        first_line = True
        temp_images = []
        temp_labels = []
        for row in csv_reader:
            if first_line:
                # print("Ignoring first line")
                first_line = False
            else:
                temp_labels.append(row[0])
                image_data = row[1:785]
                image_data_as_array = np.array_split(image_data, 28)
                temp_images.append(image_data_as_array)
        images = np.array(temp_images).astype('float')
        labels = np.array(temp_labels).astype('float')
    return images, labels

training_images, training_labels = get_data('sign_mnist_train.csv')
testing_images, testing_labels = get_data('sign_mnist_test.csv')

print(training_images.shape)
print(training_labels.shape)
print(testing_images.shape)
print(testing_labels.shape)

(27455, 28, 28)
(27455,)
(7172, 28, 28)
(7172,)

training_images = np.expand_dims(training_images, axis=3)
testing_images = np.expand_dims(testing_images, axis=3)

train_datagen = ImageDataGenerator(
    rescale=1. / 255,
    rotation_range=40,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,
    fill_mode='nearest')

validation_datagen = ImageDataGenerator(
    rescale=1. / 255)

print(training_images.shape)
print(testing_images.shape)

(27455, 28, 28, 1)
(7172, 28, 28, 1)

模型定义及训练

model = tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(64, (3, 3), activation='relu', input_shape=(28, 28, 1)),
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Conv2D(64, (3, 3), activation='relu'),
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(128, activation=tf.nn.relu),
    tf.keras.layers.Dense(26, activation=tf.nn.softmax)])

# Before modification
# model.compile(optimizer = tf.train.AdamOptimizer(),
#              loss = 'sparse_categorical_crossentropy',
#              metrics=['accuracy'])
#

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

history = model.fit_generator(train_datagen.flow(training_images, training_labels, batch_size=32),
                              steps_per_epoch=len(training_images) / 32,
                              epochs=5,
                              validation_data=validation_datagen.flow(testing_images, testing_labels, batch_size=32),
                              validation_steps=len(testing_images) / 32)

model.evaluate(testing_images, testing_labels)

Epoch 1/15
858/857 [==============================] - 24s 28ms/step - loss: 2.8346 - accuracy: 0.1390 - val_loss: 2.0034 - val_accuracy: 0.3611
Epoch 2/15
858/857 [==============================] - 24s 28ms/step - loss: 2.1808 - accuracy: 0.3113 - val_loss: 1.6024 - val_accuracy: 0.4413
Epoch 3/15
858/857 [==============================] - 24s 28ms/step - loss: 1.8357 - accuracy: 0.4080 - val_loss: 1.2396 - val_accuracy: 0.5455
Epoch 4/15
858/857 [==============================] - 23s 27ms/step - loss: 1.5816 - accuracy: 0.4843 - val_loss: 1.0176 - val_accuracy: 0.6177
Epoch 5/15
858/857 [==============================] - 23s 27ms/step - loss: 1.4129 - accuracy: 0.5408 - val_loss: 1.0217 - val_accuracy: 0.6460
Epoch 6/15
858/857 [==============================] - 24s 28ms/step - loss: 1.2887 - accuracy: 0.5829 - val_loss: 0.8568 - val_accuracy: 0.7097
Epoch 7/15
858/857 [==============================] - 23s 27ms/step - loss: 1.1790 - accuracy: 0.6110 - val_loss: 0.8046 - val_accuracy: 0.7154
Epoch 8/15
858/857 [==============================] - 23s 27ms/step - loss: 1.1048 - accuracy: 0.6346 - val_loss: 0.7046 - val_accuracy: 0.7658
Epoch 9/15
858/857 [==============================] - 24s 28ms/step - loss: 1.0506 - accuracy: 0.6494 - val_loss: 0.5894 - val_accuracy: 0.7939
Epoch 10/15
858/857 [==============================] - 24s 27ms/step - loss: 0.9842 - accuracy: 0.6732 - val_loss: 0.5881 - val_accuracy: 0.7893
Epoch 11/15
858/857 [==============================] - 23s 27ms/step - loss: 0.9247 - accuracy: 0.6924 - val_loss: 0.5618 - val_accuracy: 0.8265
Epoch 12/15
858/857 [==============================] - 23s 27ms/step - loss: 0.8841 - accuracy: 0.7070 - val_loss: 0.5434 - val_accuracy: 0.8130
Epoch 13/15
858/857 [==============================] - 24s 27ms/step - loss: 0.8422 - accuracy: 0.7228 - val_loss: 0.6367 - val_accuracy: 0.7743
Epoch 14/15
858/857 [==============================] - 23s 27ms/step - loss: 0.7952 - accuracy: 0.7332 - val_loss: 0.5824 - val_accuracy: 0.8040
Epoch 15/15
858/857 [==============================] - 24s 28ms/step - loss: 0.7751 - accuracy: 0.7463 - val_loss: 0.5417 - val_accuracy: 0.7984
225/225 [==============================] - 1s 6ms/step - loss: 309.7298 - accuracy: 0.5199

[4]:
[309.7298278808594, 0.5199386477470398]

分析

可以作图来分析一下此次训练。

本次由于数据集过小，会出现反弹现象，测试集上面的效果比训练集要好一些，当然从上面的训练过程也不难看出。

import matplotlib.pyplot as plt
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(len(acc))

plt.plot(epochs, acc, 'r', label='Training accuracy')
plt.plot(epochs, val_acc, 'b', label='Validation accuracy')
plt.title('Training and validation accuracy')
plt.legend()
plt.figure()

plt.plot(epochs, loss, 'r', label='Training Loss')
plt.plot(epochs, val_loss, 'b', label='Validation Loss')
plt.title('Training and validation loss')
plt.legend()

plt.show()

图略。