DiDong's Blog

0%

水下机器人视觉_mnist手写数字识别

Posted on In
Symbols count in article: 8.8k Reading time ≈ 8 mins.

In this notebook, we train an MLP to classify images from the MNIST database.

1. Load MNIST Database

1
2
3
4
5
6
7
from keras.datasets import mnist

# use Keras to import pre-shuffled MNIST database
(X_train, y_train), (X_test, y_test) = mnist.load_data()

print("The MNIST database has a training set of %d examples." % len(X_train))
print("The MNIST database has a test set of %d examples." % len(X_test))
The MNIST database has a training set of 60000 examples.
The MNIST database has a test set of 10000 examples.

2. Visualize the First Six Training Images

1
2
3
4
5
6
7
8
9
10
11
import matplotlib.pyplot as plt
%matplotlib inline
import matplotlib.cm as cm
import numpy as np

# plot first six training images
fig = plt.figure(figsize=(20,20))
for i in range(6):
    ax = fig.add_subplot(1, 6, i+1, xticks=[], yticks=[])
    ax.imshow(X_train[i], cmap='gray')
    ax.set_title(str(y_train[i]))

png

3. View an Image in More Detail

1
2
3
4
5
6
7
8
9
10
11
12
13
14
def visualize_input(img, ax):
    ax.imshow(img, cmap='gray')
    width, height = img.shape
    thresh = img.max()/2.5
    for x in range(width):
        for y in range(height):
            ax.annotate(str(round(img[x][y],2)), xy=(y,x),
                        horizontalalignment='center',
                        verticalalignment='center',
                        color='white' if img[x][y]<thresh else 'black')

fig = plt.figure(figsize = (12,12)) 
ax = fig.add_subplot(111)
visualize_input(X_train[0], ax)

png

4. Rescale the Images by Dividing Every Pixel in Every Image by 255

1
2
3
# rescale [0,255] --> [0,1]
X_train = X_train.astype('float32')/255
X_test = X_test.astype('float32')/255

5. Encode Categorical Integer Labels Using a One-Hot Scheme

1
2
3
4
5
6
7
8
9
10
11
12
13
from keras.utils import np_utils

# print first ten (integer-valued) training labels
print('Integer-valued labels:')
print(y_train[:10])

# one-hot encode the labels
y_train = np_utils.to_categorical(y_train, 10)
y_test = np_utils.to_categorical(y_test, 10)

# print first ten (one-hot) training labels
print('One-hot labels:')
print(y_train[:10])
Integer-valued labels:
[5 0 4 1 9 2 1 3 1 4]
One-hot labels:
[[0. 0. 0. 0. 0. 1. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.]
 [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.]
 [0. 0. 1. 0. 0. 0. 0. 0. 0. 0.]
 [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 1. 0. 0. 0. 0. 0. 0.]
 [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.]]

6. Define the Model Architecture

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten


# define the model
model = Sequential()
# 将输入展平, 不影响批量大小
model.add(Flatten(input_shape=X_train.shape[1:]))
# 全连接层
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(10, activation='softmax'))

# summarize the model
model.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten_5 (Flatten)          (None, 784)               0         
_________________________________________________________________
dense_13 (Dense)             (None, 512)               401920    
_________________________________________________________________
dropout_9 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_14 (Dense)             (None, 512)               262656    
_________________________________________________________________
dropout_10 (Dropout)         (None, 512)               0         
_________________________________________________________________
dense_15 (Dense)             (None, 10)                5130      
=================================================================
Total params: 669,706
Trainable params: 669,706
Non-trainable params: 0
_________________________________________________________________

7. Compile the Model

1
2
3
4
5
# compile the model
model.compile(loss='categorical_crossentropy', optimizer='rmsprop',
              metrics=['accuracy'])
# optimizer 优化器 如 rmsprop sgd monentum nag adagrad adadelta rmsprop
# metics 在训练和测试期间的模型评估标准, 通常使用 metircs = ['accuracy'] . 要为多输出模型的不同输出指定不同的评估标准, 还可以传递一个字典, 如 metrics = {'output_a':'accuracy'}
WARNING:tensorflow:From /home/didong/anaconda/anaconda3/lib/python3.7/site-packages/keras/optimizers.py:790: The name tf.train.Optimizer is deprecated. Please use tf.compat.v1.train.Optimizer instead.

WARNING:tensorflow:From /home/didong/anaconda/anaconda3/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:3295: The name tf.log is deprecated. Please use tf.math.log instead.

8. Calculate the Classification Accuracy on the Test Set (Before Training)

1
2
3
4
5
6
# evaluate test accuracy
score = model.evaluate(X_test, y_test, verbose=0)
accuracy = 100*score[1]

# print test accuracy
print('Test accuracy: %.4f%%' % accuracy)
Test accuracy: 7.8400%

9. Train the Model

1
2
3
4
5
6
7
8
9
10
11
12
13
from keras.callbacks import ModelCheckpoint   

# train the model
checkpointer = ModelCheckpoint(filepath='mnist.model.best.hdf5', 
                               verbose=1, save_best_only=True)
# ModelCkeckpoint 在每个训练期之后保存模型
hist = model.fit(X_train, y_train, batch_size=128, epochs=20,
          validation_split=0.2, callbacks=[checkpointer],
          verbose=1, shuffle=True)
# batch_size 每次梯度更新的样本数, 如果未指定, 默认为32
# validation_split 用作验证集的训练数据的比例
# callbacks 一系列的 keras.callbacks.Callback 实例. 一系列可以在训练时使用的回调函数
# shuffle 是否在每轮迭代之前混洗数据
WARNING:tensorflow:From /home/didong/anaconda/anaconda3/lib/python3.7/site-packages/tensorflow/python/ops/math_grad.py:1250: add_dispatch_support.<locals>.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where
Train on 48000 samples, validate on 12000 samples
Epoch 1/20
48000/48000 [==============================] - 6s 130us/step - loss: 0.2958 - acc: 0.9087 - val_loss: 0.1198 - val_acc: 0.9655

Epoch 00001: val_loss improved from inf to 0.11977, saving model to mnist.model.best.hdf5
Epoch 2/20
48000/48000 [==============================] - 5s 114us/step - loss: 0.1234 - acc: 0.9628 - val_loss: 0.1111 - val_acc: 0.9690

Epoch 00002: val_loss improved from 0.11977 to 0.11112, saving model to mnist.model.best.hdf5
Epoch 3/20
48000/48000 [==============================] - 6s 116us/step - loss: 0.0874 - acc: 0.9734 - val_loss: 0.0892 - val_acc: 0.9766

Epoch 00003: val_loss improved from 0.11112 to 0.08923, saving model to mnist.model.best.hdf5
Epoch 4/20
48000/48000 [==============================] - 6s 118us/step - loss: 0.0697 - acc: 0.9784 - val_loss: 0.0876 - val_acc: 0.9772

Epoch 00004: val_loss improved from 0.08923 to 0.08760, saving model to mnist.model.best.hdf5
Epoch 5/20
48000/48000 [==============================] - 5s 114us/step - loss: 0.0580 - acc: 0.9827 - val_loss: 0.0985 - val_acc: 0.9732

Epoch 00005: val_loss did not improve from 0.08760
Epoch 6/20
48000/48000 [==============================] - 5s 114us/step - loss: 0.0484 - acc: 0.9853 - val_loss: 0.0957 - val_acc: 0.9769

Epoch 00006: val_loss did not improve from 0.08760
Epoch 7/20
48000/48000 [==============================] - 5s 113us/step - loss: 0.0417 - acc: 0.9872 - val_loss: 0.1031 - val_acc: 0.9784

Epoch 00007: val_loss did not improve from 0.08760
Epoch 8/20
48000/48000 [==============================] - 5s 112us/step - loss: 0.0388 - acc: 0.9884 - val_loss: 0.0966 - val_acc: 0.9795

Epoch 00008: val_loss did not improve from 0.08760
Epoch 9/20
48000/48000 [==============================] - 5s 115us/step - loss: 0.0351 - acc: 0.9897 - val_loss: 0.1009 - val_acc: 0.9788

Epoch 00009: val_loss did not improve from 0.08760
Epoch 10/20
48000/48000 [==============================] - 5s 115us/step - loss: 0.0312 - acc: 0.9906 - val_loss: 0.1055 - val_acc: 0.9786

Epoch 00010: val_loss did not improve from 0.08760
Epoch 11/20
48000/48000 [==============================] - 5s 111us/step - loss: 0.0316 - acc: 0.9907 - val_loss: 0.1150 - val_acc: 0.9787

Epoch 00011: val_loss did not improve from 0.08760
Epoch 12/20
48000/48000 [==============================] - 5s 112us/step - loss: 0.0275 - acc: 0.9919 - val_loss: 0.1186 - val_acc: 0.9786

Epoch 00012: val_loss did not improve from 0.08760
Epoch 13/20
48000/48000 [==============================] - 5s 111us/step - loss: 0.0265 - acc: 0.9923 - val_loss: 0.1151 - val_acc: 0.9801

Epoch 00013: val_loss did not improve from 0.08760
Epoch 14/20
48000/48000 [==============================] - 5s 111us/step - loss: 0.0250 - acc: 0.9928 - val_loss: 0.1140 - val_acc: 0.9796

Epoch 00014: val_loss did not improve from 0.08760
Epoch 15/20
48000/48000 [==============================] - 5s 111us/step - loss: 0.0246 - acc: 0.9930 - val_loss: 0.1263 - val_acc: 0.9780

Epoch 00015: val_loss did not improve from 0.08760
Epoch 16/20
48000/48000 [==============================] - 6s 115us/step - loss: 0.0215 - acc: 0.9938 - val_loss: 0.1170 - val_acc: 0.9804

Epoch 00016: val_loss did not improve from 0.08760
Epoch 17/20
48000/48000 [==============================] - 5s 114us/step - loss: 0.0232 - acc: 0.9937 - val_loss: 0.1234 - val_acc: 0.9795

Epoch 00017: val_loss did not improve from 0.08760
Epoch 18/20
48000/48000 [==============================] - 6s 116us/step - loss: 0.0192 - acc: 0.9944 - val_loss: 0.1197 - val_acc: 0.9811

Epoch 00018: val_loss did not improve from 0.08760
Epoch 19/20
48000/48000 [==============================] - 6s 115us/step - loss: 0.0207 - acc: 0.9945 - val_loss: 0.1258 - val_acc: 0.9800

Epoch 00019: val_loss did not improve from 0.08760
Epoch 20/20
48000/48000 [==============================] - 6s 116us/step - loss: 0.0203 - acc: 0.9944 - val_loss: 0.1354 - val_acc: 0.9797

Epoch 00020: val_loss did not improve from 0.08760

10. Load the Model with the Best Classification Accuracy on the Validation Set

1
2
# load the weights that yielded the best validation accuracy
model.load_weights('mnist.model.best.hdf5')

11. Calculate the Classification Accuracy on the Test Set

1
2
3
4
5
6
# evaluate test accuracy
score = model.evaluate(X_test, y_test, verbose=0)
accuracy = 100*score[1]

# print test accuracy
print('Test accuracy: %.4f%%' % accuracy)
Test accuracy: 97.9200%
Thank you for your reward !