def test_gradients():
left = Tensor.from_builtin([[1, 2, 3], [3, 4, 1]])
right = Tensor.from_builtin([[3], [1], [-1]])
output = matrix_multiply(left, right)
gradients = output.backward(np.array([[1], [-1]]))
assert np.all(gradients[left] == [[3, 1, -1], [-3, -1, 1]])
assert np.all(gradients[right] == [[-2], [-2], [2]])
def test_tensor():
assert (
matrix_multiply(
Tensor.from_builtin([[1, 2, 3], [3, 4, 1]]),
Tensor.from_builtin([[3], [1], [-1]]),
)
== Tensor.from_builtin([[2], [12]])
)
def compute_loss():
logits = matrix_multiply(data, weights) + biases
probabilities = logits.softmax(-1)
return -(probabilities.log() * targets).sum()
def __call__(self, input: Tensor) -> Tensor:
return self.activation(
matrix_multiply(input, self.weights) + self.biases)