Всем привет, в этой статье рассмотрим простое решение задачи набора данных MNIST. Точность модели приблизительно равна 99.3%. MNIST - популярный набор данных рукописных цифр. В данной статье будем использовать разбивку тренировочные - тестовые данные из известного сайта kaggle.com. Ссылка на саму задачу:
Итак, поехали:
Import
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from keras import layers
from keras import models
from keras.utils import to_categorical
from sklearn.utils import shuffle
from sklearn.metrics import classification_report
from sklearn.model_selection import train_test_split
from skimage import img_as_ubyte
from skimage.transform import resize
from scipy.ndimage.interpolation import rotate
# train data
train_data = pd.read_csv('/kaggle/input/train.csv')
# test data
test_data = pd.read_csv('/kaggle/input/test.csv')
print('size train data:', train_data.shape)
print('size test data:', test_data.shape)
Data preparation
# split dataset train data and test data for ML
X_model = train_data.drop('label', axis=1)
y_model = train_data['label'].copy()
Y_finish = test_data
print('size train data:', X_model.shape)
print('size train labels:', y_model.shape)
print('size finish test data:', Y_finish.shape)
# frequency occurrence train labels
plt.subplots(figsize=(11,4))
y_model.value_counts().sort_index().plot('bar', color='grey')
plt.title("Frequency Histogram of Numbers in Training Data")
plt.xlabel("Number Value")
plt.ylabel("Frequency")
plt.grid(True, alpha=0.6)
# functions add modify train data
def img_rotate(df_x, angle):
""" function rotate image by angle """
change_img = np.empty([df_x.shape[0], df_x.shape[1]])
for i, image in enumerate(df_x.values):
img = rotate(image.reshape(28, 28), angle, cval=0, reshape=False, order=0)
change_img[i] = img.ravel()
return pd.DataFrame(data=change_img, columns=df_x.columns)
def img_zoom(df_x, scale):
""" function zoom image by scale """
if (scale > 0.9) or (scale < -0.9):
raise Exception('scale values must be between -0.9 and 0.9')
if scale < 0:
change_img = np.empty([df_x.shape[0], df_x.shape[1]])
for i, image in enumerate(df_x.values):
number_around = round(28*abs(scale)/2)
img = image.reshape(28, 28).astype('float32').copy()
img = np.pad(img, number_around, mode='constant')
img = resize(img, (28,28), anti_aliasing=False, mode='constant', cval=0, order=0)
change_img[i] = img.ravel().astype('int64')
return pd.DataFrame(data=change_img, columns=df_x.columns)
elif scale >= 0:
change_img = np.empty([df_x.shape[0], df_x.shape[1]])
for i, image in enumerate(df_x.values):
number_slice = round(28*abs(scale)/2)
img = image.reshape(28, 28).astype('float32').copy()
img = img[number_slice:28-number_slice, number_slice:28-number_slice]
img = resize(img, (28,28), anti_aliasing=False, mode='constant', cval=0, order=0)
change_img[i] = img.ravel().astype('int64')
return pd.DataFrame(data=change_img, columns=df_x.columns) # split data train and test and convert to Keras model
X_train, X_test, y_train, y_test = train_test_split(X_model, y_model, test_size=0.2)# add modified images training
X_train_add = X_train.append(img_zoom(X_train, 0.17))
X_train_add = X_train_add.append(img_zoom(X_train, -0.2))
X_train_add = X_train_add.append(img_rotate(X_train, 10))
X_train_add = X_train_add.append(img_rotate(X_train, -10))
# add modified images labels
y_train_add = y_train.append(y_train)
y_train_add = y_train_add.append(y_train)
y_train_add = y_train_add.append(y_train)
y_train_add = y_train_add.append(y_train)
# set train and test with add images
X_train = X_train_add.copy()
y_train = y_train_add.copy()
# shuffle full training data
X_train, y_train = shuffle(X_train, y_train)# convert train data to Keras model
X_train = X_train.values.reshape(X_train.shape[0], 28, 28 ,1)
X_train = X_train.astype('float32') / 255
X_test = X_test.values.reshape(X_test.shape[0], 28, 28, 1)
X_test = X_test.astype('float32') / 255
Y_finish = Y_finish.values.reshape(Y_finish.shape[0], 28, 28 ,1)
Y_finish = Y_finish.astype('float32') / 255# show few images
plt.figure(figsize=(18,8))
for i in range(75):
plt.subplot(5, 15, i+1)
plt.imshow(X_train[i].reshape((28,28)),cmap=plt.cm.binary)
plt.axis('off')
plt.subplots_adjust(wspace=-0.1, hspace=-0.1)
plt.show()
После преобразования видим наклоны, увеличения и уменьшения изображений. Мы увеличили стандартные 42 000 изображений до 168 000.
Deep Learning by Keras
# function build model Keras
def build_model():
# add dropout between layers
model = models.Sequential()
model.add(layers.Conv2D(32, (3,3), activation='relu', input_shape=(28,28,1)))
model.add(layers.MaxPooling2D((2,2)))
model.add(layers.Conv2D(64, (3,3), activation='relu'))
model.add(layers.MaxPooling2D((2,2)))
model.add(layers.Conv2D(64,(3,3), activation='relu'))
model.add(layers.Flatten())
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(10, activation='softmax'))
opt = optimizers.Adam(lr=0.0015, beta_1=0.9, beta_2=0.99, epsilon=None, decay=0.0, amsgrad=False)
model.compile(optimizer=opt, loss='binary_crossentropy', metrics=['accuracy'])
return model# convert train labels to categorical Keras
y_train = to_categorical(y_train, 10)
y_test = to_categorical(y_test, 10)
cnn = build_model()
cnn.fit(X_train,
y_train,
epochs=1,
batch_size=64)# get score test model
test_loss, test_acc = cnn.evaluate(X_test, y_test)
test_acc# predict test data
predict_test = cnn.predict_classes(X_test)
y_correct = np.argmax(y_test, axis=1)
correct_idx = np.nonzero(predict_test==y_correct)
incorrect_idx = np.nonzero(predict_test!=y_correct)# show report test data
target_names = ["Class {}".format(i) for i in range(10)]
print(classification_report(y_correct, predict_test, target_names=target_names))
Save results
# predict finish test data
predict = cnn.predict_classes(Y_finish)# create DataFrame predict finish data
df_out = pd.DataFrame({'ImageId': range(1, len(predict)+1),
'Label': predict})# create answers for Kaggle csv
df_out.to_csv('mnist_cnn.csv', index=False, header=True)
Количество комментариев: 0
Комментариев нет