def __call__(self, x, y_true):
""" Parameters
----------
x : mygrad.Tensor, shape=(N, C)
The C class scores for each of the N pieces of data.
y_true : Sequence[int]
The correct class-indices, in [0, C), for each datum.
Returns
-------
The average softmax loss"""
if isinstance(y_true, Tensor):
y_true = y_true.data
check_loss_inputs(x, y_true)
self.variables = (x,)
scores = x.data
log_softmax = scores - logsumexp(scores, axis=-1, keepdims=True)
label_locs = (range(len(scores)), y_true)
loss = -np.sum(log_softmax[label_locs]) / scores.shape[0]
self.back = np.exp(log_softmax)
self.back[label_locs] -= 1.0
self.back /= scores.shape[0]
return loss
def test_logsumexp(data: st.SearchStrategy, x: np.ndarray, keepdims: bool):
axes = data.draw(valid_axes(ndim=x.ndim), label="axes")
mygrad_result = logsumexp(x, axis=axes, keepdims=keepdims)
scipy_result = special.logsumexp(x, axis=axes, keepdims=keepdims)
assert_array_equal(
mygrad_result,
scipy_result,
err_msg="mygrad's implementation of logsumexp does "
"not match that of scipy's",
)