在MNIST张量流CNN上测试单个图像

Question

I've already trained a CNN based on the classic MNIST dataset, and what i'm trying to do is build a program that loads the saved model, takes an image (a handwritten digit not part of the dataset) and can predict what digit is written and output it. I've found myself stuck trying to get the code to output the guess of what digit it is(0-9).

I've already worked out how to feed in a single image in, what exactly do i write to find out what single class the model has classified the image as?

Thank you

num_channels = 1
image_size = 28
pic_root = #insert file directory here

img=mpimg.imread(pic_root)
image = img.reshape(-1,image_size,image_size, num_channels)
img = tf.cast(image, tf.float32)

with tf.Session() as session:
    saver = tf.train.import_meta_graph(save_file) #loading the saved model
    image_predict = tf.nn.softmax(img)
    print(image_predict)
    soft_max = tf.nn.softmax(logits, name="softmax_tensor")
    arg_max = tf.argmax(input=logits, axis=1)
    print(arg_max)
    print(soft_max)

image_predict, soft_max and arg_max all return something but i don't know how to get the actual prediction from this.

Answer 1

Per your code, soft_max and arg_max should be returned as arrays of length 10. The values in the array correspond to the 10 digits from 0 to 9.

Each value in soft_max array is the probability that the input image matches that particular digit. The array indexing starts at zero so the first value in array is the probability that the image is the digit 0 and the second value is the probability that the image is the digit 1 and so forth so that the tenth value is the probability that the image is 9. The predicted image is the value with the highest probability.

arg_max saves you a processing step by returning a sorted list of predicted indices (which in this case match the predicted digit). So if arg_max has [4, 9, 5, ...], the predicted digit is 4.

print(arg_max[0]) #should give you the predicted digit

Once you know that the return values are arrays as explained, the output can be decoded.