我刚从Tensorflow
开始。
我尝试实现一个模型来对MNSIT数据集中的数字进行分类。
我对Keras
很熟悉,因此我首先用它来创建模型。
Keras代码:
from keras.models import Sequential
from keras.layers import Dense
from keras.datasets import mnist
from os import path
import numpy as np
network = Sequential()
network.add(Dense(700, input_dim=784, activation='tanh'))
network.add(Dense(500, activation='tanh'))
network.add(Dense(500, activation='tanh'))
network.add(Dense(500, activation='tanh'))
network.add(Dense(10, activation='softmax'))
network.compile(loss='categorical_crossentropy', optimizer='adam')
(x_train, y_temp), (x_test, y_test) = mnist.load_data()
y_train = vectorize(y_temp) # I defined this function to create vectors of the labels. It works without issues.
x_train = x_train.reshape(x_train.shape[0], x_train.shape[1]*x_train.shape[2])
network.fit(x_train, y_train, batch_size=100, epochs=3)
x_test = x_test.reshape(x_test.shape[0], x_test.shape[1]*x_test.shape[2])
scores = network.predict(x_test)
correct_pred = 0
for i in range(len(scores)):
if np.argmax(scores[i]) == y_test[i]:
correct_pred += 1
print((correct_pred/len(scores))*100)
上面的代码为我提供了大约92%的准确性。
我尝试在Tensorflow中实现相同的模型:
import sys
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
data = input_data.read_data_sets('.', one_hot=True)
sess = tf.InteractiveSession()
x = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, 10])
w = tf.Variable(tf.zeros([784, 700]))
w2 = tf.Variable(tf.zeros([700, 500]))
w3 = tf.Variable(tf.zeros([500, 500]))
w4 = tf.Variable(tf.zeros([500, 500]))
w5 = tf.Variable(tf.zeros([500, 10]))
h1 = tf.nn.tanh(tf.matmul(x, w))
h2 = tf.nn.tanh(tf.matmul(h1, w2))
h3 = tf.nn.tanh(tf.matmul(h2, w3))
h4 = tf.nn.tanh(tf.matmul(h3, w4))
h = tf.matmul(h4, w5)
loss = tf.math.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=h, labels=y))
gradient_descent = tf.train.AdamOptimizer().minimize(loss)
correct_mask = tf.equal(tf.argmax(h, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_mask, tf.float32))
sess.run(tf.global_variables_initializer())
for i in range(3):
batch_x, batch_y = data.train.next_batch(100)
loss_print = tf.print(loss, output_stream=sys.stdout)
sess.run([gradient_descent, loss_print], feed_dict={x: batch_x, y: batch_y})
ans = sess.run(accuracy, feed_dict={x: data.test.images, y: data.test.labels})
print(ans)
但是,此代码仅使我的准确度达到了11%左右。 我尝试将纪元数增加到1000,但结果没有改变。此外,每个时期的损失是相同的(2.30)。
我在Tensorflow代码中缺少什么吗?
答案 0 :(得分:1)
结果是,问题是我将权重初始化为零!
只需更改
w = tf.Variable(tf.zeros([784, 700]))
w2 = tf.Variable(tf.zeros([700, 500]))
w3 = tf.Variable(tf.zeros([500, 500]))
w4 = tf.Variable(tf.zeros([500, 500]))
w5 = tf.Variable(tf.zeros([500, 10]))
到
w = tf.Variable(tf.random_normal([784, 700], seed=42))
w2 = tf.Variable(tf.random_normal([700, 500], seed=42))
w3 = tf.Variable(tf.random_normal([500, 500], seed=42))
w4 = tf.Variable(tf.random_normal([500, 500], seed=42))
w5 = tf.Variable(tf.random_normal([500, 10], seed=42))
进行了重大改进。