def test_on_batch(self, inputs, outputs, verbose=0):
"""
:param inputs:
:param outputs:
:param verbose:
:return:
"""
grapheme_root = Tensor(outputs[0])
vowel_diacritic = Tensor(outputs[1])
consonant_diacritic = Tensor(outputs[2])
_, grapheme_root_tru = grapheme_root.max(1)
_, vowel_diacritic_tru = vowel_diacritic.max(1)
_, consonant_diacritic_tru = consonant_diacritic.max(1)
grapheme_root_hat, vowel_diacritic_hat, consonant_diacritic_hat = self.predict(inputs)
grapheme_root_hat = grapheme_root_hat.cpu()
vowel_diacritic_hat = vowel_diacritic_hat.cpu()
consonant_diacritic_hat = consonant_diacritic_hat.cpu()
if verbose == 1:
print(grapheme_root_tru[:10])
print(grapheme_root_hat[:10])
print(vowel_diacritic_tru[:10])
print(vowel_diacritic_hat[:10])
print(consonant_diacritic_tru[:10])
print(consonant_diacritic_hat[:10])
grapheme_root = accuracy_score(grapheme_root_tru, grapheme_root_hat)
vowel_diacritic = accuracy_score(vowel_diacritic_tru, vowel_diacritic_hat)
consonant_diacritic = accuracy_score(consonant_diacritic_tru, consonant_diacritic_hat)
print("accuracy: grapheme root {}, vowel diacritic {}, consonant diacritic {}".format(grapheme_root, vowel_diacritic, consonant_diacritic))
def logsumexp(x: Tensor, dim: int) -> Tensor:
"""
Args:
x: A pytorch tensor (any dimension will do)
dim: int, over which to perform the summation.
Returns: The result of the log(sum(exp(...))) operation.
"""
xmax, _ = x.max(dim, keepdim=True)
xmax_, _ = x.max(dim)
return xmax_ + torch.log(torch.exp(x - xmax).sum(dim))
