def update_lstm(input, hiddens, cells, forget_weights, change_weights,
ingate_weights, outgate_weights):
"""One iteration of an LSTM layer."""
change = np.tanh(activations(change_weights, input, hiddens))
forget = sigmoid(activations(forget_weights, input, cells, hiddens))
ingate = sigmoid(activations(ingate_weights, input, cells, hiddens))
cells = cells * forget + ingate * change
outgate = sigmoid(activations(outgate_weights, input, cells, hiddens))
hiddens = outgate * np.tanh(cells)
return hiddens, cells
def sigmoid(x):
return 0.5*(np.tanh(x) + 1.0) # Output ranges from 0 to 1.
def sigmoid(x):
return 0.5*(np.tanh(x) + 1)
def nonlinearity(self, x):
return np.tanh(x)
def update(input, hiddens, change_weights):
return np.tanh(activations(change_weights, input, hiddens))
def predictions(W_vect, inputs):
for W, b in unpack_layers(W_vect):
outputs = np.dot(inputs, W) + b
inputs = np.tanh(outputs)
return outputs - logsumexp(outputs, axis=1, keepdims=True)
