def test_aggregation_perplexity():
# for AUC (different aggregation functionality)
metric_state = {}
metric = Perplexity()
y_pred = torch.randn(size=(100, 50))
y_true = torch.randint(0, 50, (100, ))
for yp, yt in zip(torch.split(y_pred, 10), torch.split(y_true, 10)):
metric.aggregate(metric_state, yp, yt)
gt_value = torch.exp(torch.nn.CrossEntropyLoss()(y_pred, y_true).mean()).item()
assert metric.finalize(metric_state) == approx(gt_value, NUMERIC_PRECISION)
def test_perplexity():
"""Test perplexity"""
metric_test_case(torch.tensor([[0.2, 0.8], [0.9, 0.1]]),
torch.tensor([0, 0]), Perplexity(), 2.022418975830078)