我使用Tensorflow为MNIST数据集实现了一个简单的模型。每个时期的输出平均损失:奇怪的结果
这里是模型:
X = tf.placeholder(tf.float32, [None, 784])
Y_ = tf.placeholder(tf.float32, [None, 10])
W = tf.Variable(tf.random_normal(shape=[784, 10], stddev=0.01), name="Weigths")
b = tf.Variable(tf.zeros([1,10]), name="Bias")
Y = tf.nn.softmax(tf.add(tf.matmul(X,W), b))
下面是成本函数的样子:
entropy = tf.nn.softmax_cross_entropy_with_logits(labels=Y_, logits=Y)
loss = tf.reduce_mean(entropy)
的backprop:
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(loss)
这里是我的火车循环:
for epoch in range(n_epochs):
avg_loss = 0;
n_batches = int(MNIST.train.num_examples/batch_size)
for i in range(n_batches):
X_batch, Y_batch = MNIST.train.next_batch(batch_size)
_, l, summary = sess.run([optimizer, loss, merged_summary], feed_dict={X: X_batch, Y_: Y_batch})
writer.add_summary(summary, pos)
avg_loss = l/n_batches
print('Epoch :', epoch, 'AvgLoss =', avg_loss)
print ("Accuracy:", acc.eval(feed_dict={X: MNIST.test.images, Y_: MNIST.test.labels}))
但我不明白,我对每个时期平均成本的结果:
Epoch : 0 AvgLoss = 0.0028913601962
Epoch : 1 AvgLoss = 0.00283967841755
Epoch : 2 AvgLoss = 0.0028030406345
Epoch : 3 AvgLoss = 0.002759949294
Epoch : 4 AvgLoss = 0.00283429449255
Epoch : 5 AvgLoss = 0.00276749762622
Epoch : 6 AvgLoss = 0.00276815457778
Epoch : 7 AvgLoss = 0.00279549772089
Epoch : 8 AvgLoss = 0.00277937347239
Epoch : 9 AvgLoss = 0.00274000016126
Epoch : 10 AvgLoss = 0.00275734966451
Epoch : 11 AvgLoss = 0.00278236475858
Epoch : 12 AvgLoss = 0.00275594126094
Epoch : 13 AvgLoss = 0.0027651628581
Epoch : 14 AvgLoss = 0.00275661511855
Epoch : 15 AvgLoss = 0.00275890090249
Epoch : 16 AvgLoss = 0.00273716428063
Epoch : 17 AvgLoss = 0.00273372628472
Epoch : 18 AvgLoss = 0.0027502430569
Epoch : 19 AvgLoss = 0.00279064221816
Epoch : 20 AvgLoss = 0.00273178425702
Epoch : 21 AvgLoss = 0.00277335535396
Epoch : 22 AvgLoss = 0.00276518474926
Epoch : 23 AvgLoss = 0.00276605887847
Epoch : 24 AvgLoss = 0.00275481895967
这不是减少每个循环......但它给了我一个OK精度:
Accuracy: 0.9295
任何为什么会这样想?
你可以发布你的模型代码呢? – hars
@hars编辑! –
要看损失正在减少 - 每次迭代(每批)的印刷损失不是时代。通常情况下,它在少数时代达到极小值。 – hars