Tensorflow MNIST-特定测试图像的准确性

Question

Here is the code that I used for predicting particular mnist image, how can i get accuracy by which the prediction is done ?

_pre are logits

    _x  = loaded_graph.get_tensor_by_name('x:0')
    _y  = loaded_graph.get_tensor_by_name('y:0')
    _pre  = loaded_graph.get_tensor_by_name('prediction:0')
    p = tf.argmax(_pre, 1)
    i = imageprepare('./image.png')
    print(p.eval(feed_dict={_x: [i]}))

Answer 1

I am assuming that "accuracy" means "the probability that was assigned to the selected label". From your code it is unclear how _pre was created. If it consist of the probability vectors (that is, softmax was already applied) then you can get the accuracy as follows:

acc=tf.reduce_max(_pre, 1)
print(acc.eval(feed_dict={_x: [i]}))

The reason is that p was the location of the maximal probability, while the accuracy ( acc in this case) is the maximal probability itself.

If _pre consists of the logits vectors (that is, if applying softmax on _pre will give the probability vectors),

then this will do the job:

acc=tf.reduce_max(tf.nn.softmax(_pre), 1)
print(acc.eval(feed_dict={_x: [i]}))