一,tensorflow提供了自動下載下傳mnist資料集的接口,若下載下傳不了,請嘗試翻牆後再試,或者從其他網站下載下傳。
二,訓練代碼:
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np
import os
"""------------------加載資料---------------------"""
# 載入資料
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True) #自動下載下傳mnist到MNIST_data/目錄下
trX, trY, teX, teY = mnist.train.images, mnist.train.labels, mnist.test.images, mnist.test.labels
# 改變資料格式,為了能夠輸入卷積層
trX = trX.reshape(-1, 28, 28, 1) # -1表示自适應,與batchsize相等,1表示單通道
teX = teX.reshape(-1, 28, 28, 1)
"""------------------構模組化型---------------------"""
# 定義輸入輸出的資料容器
X = tf.placeholder("float", [None, 28, 28, 1])
Y = tf.placeholder("float", [None, 10])
# 定義和初始化權重、dropout參數
def init_weights(shape):
return tf.Variable(tf.random_normal(shape, stddev=0.01))
w1 = init_weights([3, 3, 1, 32]) # 3X3的卷積核,獲得32個特征
w2 = init_weights([3, 3, 32, 64]) # 3X3的卷積核,獲得64個特征
w3 = init_weights([3, 3, 64, 128]) # 3X3的卷積核,獲得128個特征
w4 = init_weights([128 * 4 * 4, 625]) # 從卷積層到全連層
w_o = init_weights([625, 10]) # 從全連層到輸出層,輸出層的10表示是十分類問題,因為mnist是數字,識别0~9十個數字
p_keep_conv = tf.placeholder("float")
p_keep_hidden = tf.placeholder("float")
# 定義模型
def create_model(X, w1, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden):
# 第一組卷積層和pooling層
conv1 = tf.nn.conv2d(X, w1, strides=[1, 1, 1, 1], padding='SAME')
conv1_out = tf.nn.relu(conv1)
pool1 = tf.nn.max_pool(conv1_out, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
pool1_out = tf.nn.dropout(pool1, p_keep_conv)
# 第二組卷積層和pooling層
conv2 = tf.nn.conv2d(pool1_out, w2, strides=[1, 1, 1, 1], padding='SAME')
conv2_out = tf.nn.relu(conv2)
pool2 = tf.nn.max_pool(conv2_out, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
pool2_out = tf.nn.dropout(pool2, p_keep_conv)
# 第三組卷積層和pooling層
conv3 = tf.nn.conv2d(pool2_out, w3, strides=[1, 1, 1, 1], padding='SAME')
conv3_out = tf.nn.relu(conv3)
pool3 = tf.nn.max_pool(conv3_out, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
pool3 = tf.reshape(pool3, [-1, w4.get_shape().as_list()[0]]) # 轉化成一維的向量
pool3_out = tf.nn.dropout(pool3, p_keep_conv)
# 全連層
fully_layer = tf.matmul(pool3_out, w4)
fully_layer_out = tf.nn.relu(fully_layer)
fully_layer_out = tf.nn.dropout(fully_layer_out, p_keep_hidden)
# 輸出層
out = tf.matmul(fully_layer_out, w_o)
return out
model = create_model(X, w1, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden) #輸出預測機率清單
# 定義代價函數、訓練方法、預測操作
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=model, labels=Y)) #損失函數為交叉熵
train_op = tf.train.RMSPropOptimizer(0.001, 0.9).minimize(cost) #訓練操作,以最小化cost為目标
predict_op = tf.argmax(model, 1,name="predict") #預測操作
# 定義一個saver
saver=tf.train.Saver()
# 定義模型存儲路徑
ckpt_dir="./ckpt_dir"
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
"""------------------訓練模型---------------------"""
train_batch_size = 128 # 訓練集的mini_batch_size=128
test_batch_size = 256 # 測試集中調用的batch_size=256
epoches = 5 # 疊代周期
'''以上屬于計算圖,隻是單純地定義操作。而下面的表示輸入資料後,調用計算圖中定義的操作'''
with tf.Session() as sess:
"""-------訓練模型--------"""
# 初始化所有變量
tf.global_variables_initializer().run()
# 訓練操作
for i in range(epoches):
train_batch = zip(range(0, len(trX), train_batch_size),
range(train_batch_size, len(trX) + 1, train_batch_size))
for start, end in train_batch:
#執行訓練操作
sess.run(train_op, feed_dict={X: trX[start:end], Y: trY[start:end],
p_keep_conv: 0.8, p_keep_hidden: 0.5})
# 每個周期用測試集中随機抽出test_batch_size個圖檔進行測試
test_indices = np.arange(len(teX)) # 傳回一個array[0,1...len(teX)]
np.random.shuffle(test_indices) # 打亂這個array
test_indices = test_indices[0:test_batch_size]
# 擷取測試集test_batch_size章圖檔的的預測結果
predict_result = sess.run(predict_op, feed_dict={X: teX[test_indices],
p_keep_conv: 1.0,
p_keep_hidden: 1.0})
# 擷取真實的标簽值
true_labels = np.argmax(teY[test_indices], axis=1)
# 計算準确率
accuracy = np.mean(true_labels == predict_result)
print("epoch", i, ":", accuracy)
# 儲存模型
saver.save(sess,ckpt_dir+"/model.ckpt",global_step=i)
if __name__ =='__main__':
pass
運作結果:
![](https://img.laitimes.com/img/9ZDMuAjOiMmIsIjOiQnIsIyZuBnLxMTN4IzN0cTM1EzNwkTMwIzLc52YucWbp5GZzNmLn9Gbi1yZtl2Lc9CX6MHc0RHaiojIsJye.png)
并會生成模型:
每個epoch生成一個模型,每個模型分成3個檔案,五個epoch生成五個模型
三,預測代碼:
import tensorflow as tf
import numpy as np
import cv2
"""------------------構模組化型---------------------"""
# 定義輸入輸出的資料容器
X = tf.placeholder("float", [None, 28, 28, 1])
Y = tf.placeholder("float", [None, 10])
# 定義和初始化權重、dropout參數
def init_weights(shape):
return tf.Variable(tf.random_normal(shape, stddev=0.01))
w1 = init_weights([3, 3, 1, 32]) # 3X3的卷積核,獲得32個特征
w2 = init_weights([3, 3, 32, 64]) # 3X3的卷積核,獲得64個特征
w3 = init_weights([3, 3, 64, 128]) # 3X3的卷積核,獲得128個特征
w4 = init_weights([128 * 4 * 4, 625]) # 從卷積層到全連層
w_o = init_weights([625, 10]) # 從全連層到輸出層
p_keep_conv = tf.placeholder("float")
p_keep_hidden = tf.placeholder("float")
#
# # 定義模型
def create_model(X, w1, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden):
# 第一組卷積層和pooling層
conv1 = tf.nn.conv2d(X, w1, strides=[1, 1, 1, 1], padding='SAME')
conv1_out = tf.nn.relu(conv1)
pool1 = tf.nn.max_pool(conv1_out, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
pool1_out = tf.nn.dropout(pool1, p_keep_conv)
# 第二組卷積層和pooling層
conv2 = tf.nn.conv2d(pool1_out, w2, strides=[1, 1, 1, 1], padding='SAME')
conv2_out = tf.nn.relu(conv2)
pool2 = tf.nn.max_pool(conv2_out, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
pool2_out = tf.nn.dropout(pool2, p_keep_conv)
# 第三組卷積層和pooling層
conv3 = tf.nn.conv2d(pool2_out, w3, strides=[1, 1, 1, 1], padding='SAME')
conv3_out = tf.nn.relu(conv3)
pool3 = tf.nn.max_pool(conv3_out, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
pool3 = tf.reshape(pool3, [-1, w4.get_shape().as_list()[0]]) # 轉化成一維的向量
pool3_out = tf.nn.dropout(pool3, p_keep_conv)
# 全連層
fully_layer = tf.matmul(pool3_out, w4)
fully_layer_out = tf.nn.relu(fully_layer)
fully_layer_out = tf.nn.dropout(fully_layer_out, p_keep_hidden)
# 輸出層
out = tf.matmul(fully_layer_out, w_o)
return out
'''上面是與訓練代碼一樣的模型代碼'''
model = create_model(X, w1, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden)
predict_op = tf.argmax(model, 1,name="predict") #預測操作
# 定義一個saver
saver=tf.train.Saver()
with tf.Session() as sess:
"""-------訓練模型--------"""
# 初始化所有變量
tf.global_variables_initializer().run()
"""-----加載模型,用導入的圖檔進行測試--------"""
# 載入圖檔
src = cv2.imread('./7.png')
cv2.imshow("待測圖檔", src)
# 将圖檔轉化為28*28的灰階圖
src = cv2.cvtColor(src, cv2.COLOR_BGR2GRAY)
print(src.shape)
dst = cv2.resize(src, (28, 28), interpolation=cv2.INTER_CUBIC)
# # 将灰階圖歸一化
picture = np.zeros((28, 28))
for i in range(0, 28):
for j in range(0, 28):
picture[i][j] = (255 - dst[i][j])/255
picture = picture.reshape(1, 28, 28, 1) #由于網絡輸入格式是4維的,是以輸入的圖檔也要轉成(batchsize,h,w,c)四維
# 載入模型
saver.restore(sess,"./ckpt_dir/model.ckpt-4")
# 進行預測
predict_result = sess.run(predict_op, feed_dict={X: picture,
p_keep_conv: 1.0,
p_keep_hidden: 1.0})
print("你導入的圖檔是:",predict_result[0])
cv2.waitKey(0)
預測結果:
先用畫圖工具畫出一張32x32的手寫體:
再輸入到預測代碼中運作: