Keras密集层输出-形状错误

Question

I was working on an NLP problem with LSTM. The problem is a multiclass classification with 3 classes (1,2 & 3). So I converted the target classes with the following code: y_train=to_catgorical(y_train) and similarly for y_test . But in the output dense layer when I write model.add((Dense(3,activation='softmax')) , I get the following error:

Error when checking target: expected dense_1 to have shape (None, 3) but got array with shape (658118, 4)

However when I fix it to model.add((Dense(4,activation='softmax')) , ie n+1 (n=no. of classes to be predicted) it works. But in Keras examples they have worked with cifar10 dataset and have used number of classes as 10, not 11 and it works. Also in case of binary classification (2 classes to be predicted) we use just 1 ouptut, ie model.add(Dense(1,"sigmoid")) . I've gone through this post but still not been able to find a convincing logic for this thing, so I want to get this concept cleared with the Dense layer output shape in Keras.

PS What I've understood is, Keras considers classes from 0:num_classes , so we have to give one more transformation. But to ask one thing that '0' vector layer then has to remain unused right? And in case of cifar10 dataset, is it that the classes are 0:9 , that's why num_classes=10 work? If that is the case then if we have to make a prediction of 3 classes say (0,1 & 2) then we can use num_classes=3 right?

Answer 1

Perhaps it is important to note the use of Keras framework to provide an one hot encoding function:

from keras.datasets import cifar10
from keras.utils import np_utils    
y_label_train_OneHot = np_utils.to_categorical(y_label_train)
y_label_test_OneHot = np_utils.to_categorical(y_label_test)

You can see this code code In short, I personally think that the use of different functions lead to a different set of categories.