def test_compute_rank_from_scores(self):
"""Test the _compute_rank_from_scores() function."""
batch_size = 3
all_scores = torch.tensor([
[2., 2., 1., 3., 5.],
[1., 1., 3., 4., 0.],
[1., 1., 3., float('nan'), 0],
])
# true_score: (2, 3, 3)
true_score = torch.tensor([2., 3., 3.]).view(batch_size, 1)
exp_best_rank = torch.tensor([3., 2., 1.])
exp_worst_rank = torch.tensor([4., 2., 1.])
exp_avg_rank = 0.5 * (exp_best_rank + exp_worst_rank)
exp_adj_rank = exp_avg_rank / torch.tensor([(5 + 1) / 2, (5 + 1) / 2,
(4 + 1) / 2])
ranks = compute_rank_from_scores(true_score=true_score,
all_scores=all_scores)
best_rank = ranks.get('best')
assert best_rank.shape == (batch_size, )
assert (best_rank == exp_best_rank).all()
worst_rank = ranks.get('worst')
assert worst_rank.shape == (batch_size, )
assert (worst_rank == exp_worst_rank).all()
avg_rank = ranks.get('avg')
assert avg_rank.shape == (batch_size, )
assert (avg_rank == exp_avg_rank).all(), (avg_rank, exp_avg_rank)
adj_rank = ranks.get('adj')
assert adj_rank.shape == (batch_size, )
assert (adj_rank == exp_adj_rank).all(), (adj_rank, exp_adj_rank)
def test_compute_rank_from_scores(self):
"""Test the _compute_rank_from_scores() function."""
batch_size = 3
all_scores = torch.tensor([
[2., 2., 1., 3., 5.],
[1., 1., 3., 4., 0.],
[1., 1., 3., float('nan'), 0],
])
# true_score: (2, 3, 3)
true_score = torch.as_tensor([2., 3., 3.]).view(batch_size, 1)
exp_best_rank = torch.as_tensor([3., 2., 1.])
exp_worst_rank = torch.as_tensor([4., 2., 1.])
exp_avg_rank = 0.5 * (exp_best_rank + exp_worst_rank)
exp_exp_rank = torch.as_tensor([(5 + 1) / 2, (5 + 1) / 2, (4 + 1) / 2])
ranks = compute_rank_from_scores(true_score=true_score,
all_scores=all_scores)
optimistic_rank = ranks.get(RANK_OPTIMISTIC)
assert optimistic_rank.shape == (batch_size, )
assert (optimistic_rank == exp_best_rank).all()
pessimistic_rank = ranks.get(RANK_PESSIMISTIC)
assert pessimistic_rank.shape == (batch_size, )
assert (pessimistic_rank == exp_worst_rank).all()
realistic_rank = ranks.get(RANK_REALISTIC)
assert realistic_rank.shape == (batch_size, )
assert (realistic_rank == exp_avg_rank).all(), (realistic_rank,
exp_avg_rank)
expected_realistic_rank = ranks.get(RANK_EXPECTED_REALISTIC)
assert expected_realistic_rank is not None
assert expected_realistic_rank.shape == (batch_size, )
assert (expected_realistic_rank == exp_exp_rank).all(), (
expected_realistic_rank, exp_exp_rank)
