def test_perfectly_predicted_samples(torch_device, identity_model):
myThresholder = pseudo.Threshold([.7, .8]).to(
torch_device) # thresholds for a 2 class problem.
# corresponding dists.---- [1.0,0.0] [0, 1.0] [1.0 , 0] [ 1.0, 0]
predictions = torch.Tensor([[1000, 0], [0, 1000], [1000.0, 0.0],
[1000.0, 0]]).to(torch_device)
targets = torch.Tensor([0, 1, -1, -1]).long().to(torch_device)
myPseudo = pseudo.PL(identity_model, myThresholder).to(torch_device)
expected = torch.Tensor(0).to(torch_device)
assert torch.allclose(myPseudo.get_technique_cost(predictions, targets),
expected)
def test_pseudo_one_sample_supervised(torch_device, identity_model):
myThresholder = pseudo.Threshold([.7, .8]).to(
torch_device) # thresholds for a 2 class problem.
# corresponding dists.----[.73, .26]
predictions = torch.Tensor([[96, 95]]).to(torch_device)
targets = torch.Tensor([0]).long().to(torch_device)
myPseudo = pseudo.PL(identity_model, myThresholder,
ssml_mode=False).to(torch_device)
expected = torch.Tensor([0.3133]).to(torch_device)
assert torch.allclose(myPseudo.get_technique_cost(predictions, targets),
expected,
atol=1e-2)
def test_pseudo_semi_sup(torch_device, identity_model):
myThresholder = pseudo.Threshold([.7, .8]).to(
torch_device) # thresholds for a 2 class problem.
# corresponding dists.---- [1.0,0.0] [.95, .05][0, 1.0] [.12,.88] [.27,.73] [.5, .5]
predictions = torch.Tensor([[90, 10], [90, 87], [30, 70], [88, 90],
[95, 96], [50, 50]]).to(torch_device)
targets = torch.Tensor([0, -1, 1, -1, 0, -1]).long().to(torch_device)
myPseudo = pseudo.PL(identity_model, myThresholder).to(torch_device)
expected = torch.Tensor([0.088]).to(torch_device)
assert torch.allclose(myPseudo.get_technique_cost(predictions, targets),
expected,
atol=1e-2)
def test_all_zero_weights_semi_sup(torch_device, identity_model):
myThresholder = pseudo.Threshold([.7, .8]).to(
torch_device) # thresholds for a 2 class problem.
# corresponding dists.---- [.62, .38] [.27,.73] [.5, .5]
predictions = torch.Tensor([[95.5, 95], [95, 96], [50,
50]]).to(torch_device)
targets = torch.Tensor([-1, -1, -1]).long().to(torch_device)
myPseudo = pseudo.PL(identity_model, myThresholder).to(torch_device)
expected = torch.Tensor(0).to(torch_device)
assert torch.allclose(myPseudo.get_technique_cost(predictions, targets),
expected,
atol=1e-2)
def test_pseudo_one_sample_semi_sup(torch_device, identity_model):
myThresholder = pseudo.Threshold([.7, .8]).to(
torch_device) # thresholds for a 2 class problem.
# corresponding dists.----[.73, .26]
predictions = torch.Tensor([[96, 95]]).to(torch_device)
targets = torch.Tensor([-1]).long().to(torch_device)
myPseudo = pseudo.PL(identity_model, myThresholder.to(torch_device))
# np.testing.assert_almost_equal(myPseudo(predictions, targets), 0.3133, decimal=3)
expected = torch.Tensor([0.3133]).to(torch_device)
assert torch.allclose(myPseudo.get_technique_cost(predictions, targets),
expected,
atol=1e-2)
def test_pseudo_supervised(torch_device, identity_model):
myThresholder = pseudo.Threshold([.7, .8]).to(
torch_device) # thresholds for a 2 class problem.
# corresponding dists.---- [1.0,0.0] [1.0,0] [0 , 1.0] [.27, .73] [.27,.73] [.5, .5]
predictions = torch.Tensor([[90, 10], [90, 10], [30, 70], [96, 95],
[95, 96], [50, 50]]).to(torch_device)
targets = torch.Tensor([1, 0, 1, 1, 1, 1]).long().to(torch_device)
myPseudo = pseudo.PL(identity_model, myThresholder,
ssml_mode=False).to(torch_device)
expected = torch.Tensor([20.328]).to(torch_device)
assert torch.allclose(myPseudo.get_technique_cost(predictions, targets),
expected,
atol=1e-2)
def test_multiclass_samples(torch_device, identity_model):
myThresholder = pseudo.Threshold([.7, .8, .99]).to(
torch_device) # thresholds for a 3 class problem.
# corresponding dists.--[.33,.33,.33] [~0, ~0.27, .73] [.88, .12, ~0]
predictions = torch.Tensor([[1, 1, 1], [1, 10, 11], [12, 10, 1], [1, 1, 1],
[1, 10, 11], [12, 10, 1]]).to(torch_device)
targets = torch.Tensor([0, 1, 2, -1, -1, -1]).long().to(torch_device)
# This should only count the CE loss to the last sample
myPseudo = pseudo.PL(identity_model, myThresholder).to(torch_device)
expected = torch.Tensor([0.1269]).to(torch_device)
assert torch.allclose(myPseudo.get_technique_cost(predictions, targets),
expected,
atol=1e-2)