def forwards(self, weights: np.ndarray) -> np.ndarray:
confirm_shape(weights, (3073, 10))
# Each of the 10 weight matrices represents a class.
# We calculate a score for each class by multiplying the
# input against the weight matrix for that class.
output = weights * self.input
confirm_shape(output, (10))
self.output = output
return output
def forwards(self, weights: np.ndarray) -> np.ndarray:
confirm_shape(weights, (3073, 10))
# Each of the 10 weight matrices represents a class.
# For each input we calculate a score for each class by multiplying
# the input against the weight matrix for that class.
output = np.dot(self.inputs, weights)
# We should end up with 10 class scores for each input.
confirm_shape(output, (500, 10))
self.output = output
return output
def backwards(self, gradient) -> np.ndarray:
confirm_shape(gradient, (500, 10))
# First we want to calculate the gradient for each weight value
# with respect to each input. So: how much did that
# particular weight value effect the score for that input?
all_input_grads = gradient[:, :, np.newaxis] * self.inputs[:,
np.newaxis]
# Then, to get an overall gradient, i.e. how much did each weight
# effect the scores overall.
# We get this just by summing all the values for each weight over
# all the inputs.
output = np.sum(all_input_grads, axis=0)
confirm_shape(output, (10, 3073))
return output
def __init__(self, inputs):
Node.__init__(self)
self.inputs = inputs
confirm_shape(inputs, (500, 3073))
def __init__(self, input):
Node.__init__(self)
self.input = input
confirm_shape(inputs, (3073))