class SoftmaxPair(ActivationPair):
neon_activation = Softmax()
tolerance = 1e-6
def reference_value(self, x):
return (np.exp(x - 1) / np.sum(np.exp(x - 1), axis=0, keepdims=True))
def reference_derivative(self, x):
f = self.reference_value(x)
return f * (1.0 - f)
def test_softmax_big_inputs(transformer_factory):
"""
This fails with memory error because the ex.derivative function
attempts to compute the full derivative.
Keeping this test since it was in original neon.
"""
inputs = np.random.random((1000, 128))
outputs = (np.exp(inputs - 1) / np.sum(np.exp(inputs - 1)))
outputs = outputs * (1 - outputs) # shortcut only
compare_tensors(Softmax(), inputs, outputs, deriv=True, tol=1e-6)
def test_softmax_derivative(transformer_factory):
inputs = np.array([0, 1, -2], dtype=np.float).reshape((3, 1))
outputs = (np.exp(inputs - 1) / np.sum(np.exp(inputs - 1)))
outputs = outputs * (1 - outputs) # shortcut only
compare_tensors(Softmax(), inputs, outputs, deriv=True, tol=1e-6)
def test_softmax(transformer_factory):
inputs = np.array([0, 1, -2]).reshape((3, 1))
outputs = (np.exp(inputs - 1) / np.sum(np.exp(inputs - 1))).reshape((3, 1))
compare_tensors(Softmax(), inputs, outputs, tol=1e-5)