我已经训练了一个卷积神经网络(CNN),其中包含我在二进制文件中的以下数据(标签,文件名,数据(像素)):
[array([2, 1, 0, 2, 1, 0, 2, 1, 0, 2, 1, 0, 2, 1, 0, 2, 1, 0, 2, 1, 0, 2, 1,
0, 2, 1, 0, 2, 1, 0]), array(['10_c.jpg', '10_m.jpg', '10_n.jpg', '1_c.jpg',
'1_m.jpg', '1_n.jpg', '2_c.jpg', '2_m.jpg',
'2_n.jpg', '3_c.jpg', '3_m.jpg', '3_n.jpg',
'4_c.jpg', '4_m.jpg', '4_n.jpg', '5_c.jpg',
'5_m.jpg', '5_n.jpg', '6_c.jpg', '6_m.jpg',
'6_n.jpg', '7_c.jpg', '7_m.jpg', '7_n.jpg',
'8_c.jpg', '8_m.jpg', '8_n.jpg', '9_c.jpg',
'9_m.jpg', '9_n.jpg'],
dtype='<U15'), array([[255, 252, 255, ..., 255, 255, 255],
[136, 137, 138, ..., 114, 110, 111],
[200, 200, 199, ..., 179, 178, 177],
...,
[146, 157, 165, ..., 202, 202, 201],
[228, 225, 222, ..., 219, 221, 223],
[128, 127, 127, ..., 133, 129, 127]])]
每批包含所有图像,并运行30个epoh:
EPOCH 0
0 0.476923
DONE WITH EPOCH
EPOCH 1
0 0.615385
DONE WITH EPOCH
EPOCH 2
0 0.615385
DONE WITH EPOCH
EPOCH 3
0 0.538462
DONE WITH EPOCH
EPOCH 4
0 0.384615
DONE WITH EPOCH
...
...
EPOCH 28
0 0.615385
DONE WITH EPOCH
EPOCH 29
0 0.692308
DONE WITH EPOCH
我的问题是我想尝试新的图像(测试),并想知道返回的类(0,1,2)。在这种情况下我该怎么办?换句话说,我训练了CNN,但是如何测试呢?
修改-1
对于评估准确性点,我在测试20张图片时得到以下结果:
EPOCH 0
0 1.0
DONE WITH EPOCH
EPOCH 1
0 1.0
DONE WITH EPOCH
EPOCH 2
0 1.0
DONE WITH EPOCH
EPOCH 3
0 1.0
DONE WITH EPOCH
EPOCH 4
0 1.0
DONE WITH EPOCH
EPOCH 5
0 1.0
DONE WITH EPOCH
EPOCH 6
0 1.0
DONE WITH EPOCH
EPOCH 7
0 1.0
DONE WITH EPOCH
EPOCH 8
0 1.0
DONE WITH EPOCH
EPOCH 9
0 1.0
DONE WITH EPOCH
EPOCH 10
0 1.0
DONE WITH EPOCH
EPOCH 11
0 1.0
DONE WITH EPOCH
EPOCH 12
0 1.0
DONE WITH EPOCH
EPOCH 13
0 1.0
DONE WITH EPOCH
EPOCH 14
0 1.0
DONE WITH EPOCH
EPOCH 15
0 1.0
DONE WITH EPOCH
EPOCH 16
0 1.0
DONE WITH EPOCH
EPOCH 17
0 1.0
DONE WITH EPOCH
EPOCH 18
0 1.0
DONE WITH EPOCH
EPOCH 19
0 1.0
DONE WITH EPOCH
EPOCH 20
0 1.0
DONE WITH EPOCH
EPOCH 21
0 1.0
DONE WITH EPOCH
EPOCH 22
0 1.0
DONE WITH EPOCH
EPOCH 23
0 1.0
DONE WITH EPOCH
EPOCH 24
0 1.0
DONE WITH EPOCH
EPOCH 25
0 1.0
DONE WITH EPOCH
EPOCH 26
0 1.0
DONE WITH EPOCH
EPOCH 27
0 1.0
DONE WITH EPOCH
EPOCH 28
0 1.0
DONE WITH EPOCH
EPOCH 29
0 1.0
DONE WITH EPOCH
当应用获取网络为测试数据生成的标签时,我得到以下内容:
EPOCH 0
0 0.0
DONE WITH EPOCH
EPOCH 1
0 0.0
DONE WITH EPOCH
EPOCH 2
0 0.0
DONE WITH EPOCH
EPOCH 3
0 0.0
DONE WITH EPOCH
EPOCH 4
0 0.0
DONE WITH EPOCH
EPOCH 5
0 0.0
DONE WITH EPOCH
EPOCH 6
0 0.0
DONE WITH EPOCH
EPOCH 7
0 0.0
DONE WITH EPOCH
EPOCH 8
0 0.0
DONE WITH EPOCH
EPOCH 9
0 0.0
DONE WITH EPOCH
EPOCH 10
0 0.0
DONE WITH EPOCH
EPOCH 11
0 0.0
DONE WITH EPOCH
EPOCH 12
0 0.0
DONE WITH EPOCH
EPOCH 13
0 0.0
DONE WITH EPOCH
EPOCH 14
0 0.0
DONE WITH EPOCH
EPOCH 15
0 0.0
DONE WITH EPOCH
EPOCH 16
0 0.0
DONE WITH EPOCH
EPOCH 17
0 0.0
DONE WITH EPOCH
EPOCH 18
0 0.0
DONE WITH EPOCH
EPOCH 19
0 0.0
DONE WITH EPOCH
EPOCH 20
0 0.0
DONE WITH EPOCH
EPOCH 21
0 0.0
DONE WITH EPOCH
EPOCH 22
0 0.0
DONE WITH EPOCH
EPOCH 23
0 0.0
DONE WITH EPOCH
EPOCH 24
0 0.0
DONE WITH EPOCH
EPOCH 25
0 0.0
DONE WITH EPOCH
EPOCH 26
0 0.0
DONE WITH EPOCH
EPOCH 27
0 0.0
DONE WITH EPOCH
EPOCH 28
0 0.0
DONE WITH EPOCH
EPOCH 29
0 0.0
DONE WITH EPOCH
为什么我要么0
还是1
?这些值是否有意义(即没有分数)?
修改-2
对于获取网络为测试数据生成的标签,在打印出标签值和每个纪元的准确性时,我得到以下内容(标签总是0
,尽管我我只期待0
或2
,准确度为1
):
EPOCH 0
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] 1.0
DONE WITH EPOCH
EPOCH 1
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] 1.0
DONE WITH EPOCH
EPOCH 2
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] 1.0
DONE WITH EPOCH
EPOCH 3
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] 1.0
DONE WITH EPOCH
EPOCH 4
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] 1.0
DONE WITH EPOCH
EPOCH 5
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] 1.0
DONE WITH EPOCH
.....
.....
EPOCH 28
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] 1.0
DONE WITH EPOCH
EPOCH 29
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] 1.0
DONE WITH EPOCH
感谢。
答案 0 :(得分:7)
一般性讨论
通常,为了测试神经网络,您需要获取未用于训练的新标记数据,在此数据上应用网络(即应用前馈过程),并评估准确性结果(与您知道的标签相比是真实的)。
如果您没有这样的新数据(也就是说,如果您使用了所有数据进行培训)并且无法生成新数据,我建议您将您的培训数据与培训和测试分开,并重新运行您从一开始就训练数据的训练程序。重要的是,测试数据将是未使用的数据,以便能够评估模型的性能。
评估准确性
现在,假设您正在讨论来自this问题的网络, 您可以执行类似的操作来衡量测试数据的准确性:
fathers_occupation
其中accuracy_test = sess.run(accuracy, feed_dict={x: test_data, y: test_onehot_vals})
和test_data
是您的测试图片(和相应的标签)。
回想一下,对于培训,您可以执行以下操作:
test_onehot_vals
请注意,我在评估_, accuracy_val = sess.run([train_op, accuracy], feed_dict={x: batch_data, y: batch_onehot_vals})
时没有使用train_op
。这是因为当您测试性能时,您不会优化权重或类似的任何内容(accuracy_test
执行此操作)。您只需应用当前拥有的网络。
获取网络为测试数据生成的标签
最后,如果您想要测试数据的实际标签,则需要获得train_op
的值。因此,您可以将其设置为单独的变量,例如在行
tf.argmax(model_op, 1)
你可以这样做:
correct_pred = tf.equal(tf.argmax(model_op, 1), tf.argmax(y,1))
然后将其与res_model=tf.argmax(model_op, 1)
correct_pred = tf.equal(res_model, tf.argmax(y,1))
一起评估,如下所示:
accuracy_test
在未标记数据上应用网络
完成网络测试后,假设您对结果感到满意,您可以继续并在新的和未标记的数据上应用网络。例如,通过
res, accuracy_test = sess.run([res_model,accuracy], feed_dict={x: test_data, y: test_onehot_vals}).
。
请注意,为了生成res_new = sess.run(res_model, feed_dict={x: new_data})
(这基本上意味着只在输入上应用网络),您不需要任何标签,因此Feed_dict中不需要res_model
个值。 y
将是新标签。