In the week-2 assignment, I came across a statement in the notebook that “Using Keras’ Tokenizer yields values that start at 1 rather than at 0. and asks me to subtract 1 from each value in the output array”. Why am I doing this? The number of values in the output array (label_seq_np) is equal to 1780, from this I understand that there will be 1780 neurons present in out first layer input layer. To work our way around, why should we add one extra neuron in out architecture?
Numeric labels produced by keras tokenizer starts from 1. This is because, 0 is reserved for internal use.
A multiclass classification model has number of output units set to the number of classes.Here’s what happens during a single forward pass:
- Feed input to the model
- For each layer:
a. Generate output by first performing the linear transformation ofdotProduct(weights, input) + bias
b. Apply layer activation function to the output of previous step.
c. Feed this output to the next layer - In the final layer, once you apply the correct activation function, the outputs represent probability of the user input belonging to that particular class. The prediction for an input is the
argmaxof the output generated by the output layer.
The prediction is going to be in range [0, num_num_classes - 1]. This explains the -1 when using labels. You can add another neuron but it’s not required. This becomes apparent when doing a binary classification. The model becomes slow due to 2 output units instead of 1.