Handwritten number write recognition, I believe, is to build a model, the input data, use optimization algorithm to minimize the loss value, w, so as to update the network, the cycle to loss value changes slowly, then training the model,
But the picture code, builds the model of the final is a full connection layer, the output of 10 data, but there is no softmax into probability, nor into 0-9 categories, so, the training model, loss value is how to calculate? When using a test set, loss value and accuracy how come, because I think this data can not with labels, calculated the loss value, accuracy,