def predict(params, inputs):
activations = inputs
for w, b in params[:-1]:
outputs = np.dot(activations, w) + b
activations = np.tanh(outputs)
final_w, final_b = params[-1]
logits = np.dot(activations, final_w) + final_b
return logits - logsumexp(logits, axis=1, keepdims=True)
def predict(params, inputs):
for w, b in params:
outputs = np.dot(inputs, w) + b
inputs = np.tanh(outputs)
return outputs - logsumexp(outputs, axis=1, keepdims=True)
def logsoftmax(x, axis=-1):
"""Apply log softmax to an array of logits, log-normalizing along an axis."""
return x - logsumexp(x, axis, keepdims=True)
def lax_fun(array_to_reduce):
return lsp_misc.logsumexp(array_to_reduce, axis, keepdims=keepdims)
def one_sided_exp(w_F):
DeltaF = -(logsumexp(-w_F) - np.log(len(w_F)))
return DeltaF
