def test_givenAPerfectPredictionTensor_whenNLLLossPerTag_thenLossIs0(self):
# Tags prediction value are 'inverted' to mimic the log
first_token_first_element_of_the_batch = [
0., 1.
] # the predicted token is the first class
first_token_second_element_of_the_batch = [
1., 0.
] # the predicted token is the second class
second_token_first_element_of_the_batch = [
1., 0.
] # the predicted token is the second class
second_token_second_element_of_the_batch = [
0., 1.
] # the predicted token is the first class
predict_tensor = torch.tensor(
[[
first_token_first_element_of_the_batch,
first_token_second_element_of_the_batch
],
[
second_token_first_element_of_the_batch,
second_token_second_element_of_the_batch
]],
device=self.a_device)
actual = nll_loss(predict_tensor,
self.a_short_ground_truth).detach().tolist()
expected = torch.tensor(1, device=self.a_device).tolist()
self.assertAlmostEqual(expected, actual, delta=5)
def test_givenACompletelyWrongPredictionTensor_whenNLLLossPerTag_thenLossIsMinus2(
self, ):
# Tags prediction value are "inverted" to mimic the log
first_token_first_element_of_the_batch = [
1.0,
0.0,
] # the predicted token is the second class
first_token_second_element_of_the_batch = [
0.0,
1.0,
] # the predicted token is the first class
second_token_first_element_of_the_batch = [
0.0,
1.0,
] # the predicted token is the first class
second_token_second_element_of_the_batch = [
1.0,
0.0,
] # the predicted token is the second class
predict_tensor = torch.tensor(
[
[
first_token_first_element_of_the_batch,
first_token_second_element_of_the_batch,
],
[
second_token_first_element_of_the_batch,
second_token_second_element_of_the_batch,
],
],
device=self.a_device,
)
actual = nll_loss(predict_tensor,
self.a_short_ground_truth).detach().tolist()
expected = torch.tensor(0.0, device=self.a_device)
self.assertAlmostEqual(expected, actual, delta=5)
def test_givenAPredictionTensor_whenNLLLossPerTag_thenLossIsOk(self):
# need to convert to list and use float since not working almost equal for tensor
actual = nll_loss(self.a_prediction_tensor,
self.ground_truth).detach().tolist()
expected = self.a_loss.tolist()
self.assertAlmostEqual(expected, actual, delta=5)