Decide the right type of dot product depending on the input arguments

@param current: index of current step @param values: list of tuples (name, value_for_last_step). The progress bar

Build the Encoder Computational Graph

Numerically stable log(sum(exp(A))). Can also be used in softmax function. c is the additional input when it doesn't require masking

Build the Decoder Computational Graph For training/testing

Build the Encoder Computational Graph

=================================================== Grid (2D as an example) ------------- | c' d' | Grid B

Build the Encoder Computational Graph

Build a sampler which only steps once. Typically it only works for one word a time?

Consider the situation that gradient is weighted.

What's going on here? [1] change the 2D matrix into 3D. [2]

Build the model weights

Build the Decoder Computational Graph

Build a sampler which only steps once. Typically it only works for one word a time?

Return the node for key t if is in the tree, or None otherwise.

Insert key t into this BST, modifying it in-place.

Build the Decoder Computational Graph For training/testing

Build a sampler which only steps once. Typically it only works for one word a time?

build the hidden action prediction.

build the sampler function here <:::>

Build the Pointer Network Decoder Computational Graph

LSTM [h', m'] = LSTM(x, h, m): gi = sigmoid(Wi * x + Ui * h + Vi * m) # Vi is peep-hole gf = sigm