TensorFlow - 训练准确性没有改善MNIST数据

Question

我编写了一个带有tensorflow的程序来处理Kaggle的数字识别器问题。程序可以正常运行，但训练精度总是很低，约为10％，如下所示：

step 0, training accuracy 0.11
step 100, training accuracy 0.13
step 200, training accuracy 0.21
step 300, training accuracy 0.12
step 400, training accuracy 0.07
step 500, training accuracy 0.08
step 600, training accuracy 0.15
step 700, training accuracy 0.05
step 800, training accuracy 0.08
step 900, training accuracy 0.12
step 1000, training accuracy 0.05
step 1100, training accuracy 0.09
step 1200, training accuracy 0.12
step 1300, training accuracy 0.1
step 1400, training accuracy 0.08
step 1500, training accuracy 0.11
step 1600, training accuracy 0.17
step 1700, training accuracy 0.13
step 1800, training accuracy 0.11
step 1900, training accuracy 0.13
step 2000, training accuracy 0.07
……

以下是我的代码：

def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)

def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)

def conv2d(x, w):
    return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
    # ksize = [batch, heigh, width, channels], strides=[batch, stride, stride, channels]
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

x = tf.placeholder(tf.float32, [None, 784])      
y_ = tf.placeholder(tf.float32, [None, 10])      
keep_prob = tf.placeholder(tf.float32)

x_image = tf.placeholder(tf.float32, [None, 28, 28, 1])

w_conv1 = weight_variable([5, 5, 1, 32])    
b_conv1 = bias_variable([32])
h_conv1 = tf.nn.relu(conv2d(x_image, w_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

w_conv2 = weight_variable([5, 5, 32, 64])    
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, w_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

w_fc1 = weight_variable([7 * 7 * 64, 1024])   
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])     
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, w_fc1) + b_fc1)
# dropout
keep_prob = tf.placeholder(tf.float32)      
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# softmax
w_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, w_fc2) + b_fc2)

cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv), reduction_indices=[1]))
train_step = tf.train.AdamOptimizer(10e-4).minimize(cross_entropy)

correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

def get_batch(i, size, train, label):
    startIndex = (i * size) % 42000
    endIndex = startIndex + size
    batch_X = train[startIndex : endIndex]
    batch_Y = label[startIndex : endIndex]
    return batch_X, batch_Y


data = pd.read_csv('train.csv')
train_data = data.drop(['label'], axis=1)
train_data = train_data.values.astype(dtype=np.float32)
train_data = train_data.reshape(42000, 28, 28, 1)

label_data = data['label'].tolist()
label_data = tf.one_hot(label_data, depth=10)
label_data = tf.Session().run(label_data).astype(dtype=np.float64)


batch_size = 100                             
tf.global_variables_initializer().run()

for i in range(20000):   
    batch_x, batch_y = get_batch(i, batch_size, train_data, label_data)
    if i % 100 == 0:
        train_accuracy = accuracy.eval(feed_dict={x_image: batch_x, y_: batch_y, keep_prob: 1.0})
        print("step %d, training accuracy %g" % (i, train_accuracy))
    train_step.run(feed_dict={x_image: batch_x, y_: batch_y, keep_prob: 0.9})

我不知道我的节目有什么问题。

Answer 1

我建议你改变你的bias_variable函数 - 不确定tf.Variable(tf.constant)的行为，加上我们通常将偏差初始化为零，而不是0.1：

def bias_variable(shape):
    return tf.zeros((shape), dtype = tf.float32)

如果这没有帮助，请尝试使用stddev=0.01初始化您的权重