Log In Sign Up
Models:
Renee Dark
ReneeD/
torch_ecg
0
Downloads: 1
[b48499]: / test / test_models / test_loss.py

Download this file

154 lines (127 with data), 7.0 kB 

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
"""

"""



import pytest

import torch



from torch_ecg.models.loss import (

    AsymmetricLoss,

    BCEWithLogitsWithClassWeightLoss,

    FocalLoss,

    MaskedBCEWithLogitsLoss,

    WeightedBCELoss,

    setup_criterion,

)



DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")





inp = torch.tensor([[10.0, -10.0], [-10.0, 10.0], [-10.0, 10.0]]).to(DEVICE)

targ_1 = torch.tensor([[1.0, 0.0], [0.0, 1.0], [0.0, 1.0]]).to(DEVICE)

targ_0 = torch.tensor([[0.0, 1.0], [1.0, 0.0], [1.0, 0.0]]).to(DEVICE)

targ_mixed = torch.tensor([[1.0, 0.0], [0.0, 1.0], [1.0, 0.0]]).to(DEVICE)

targ_1_soft = torch.tensor([[0.9, 0.1], [0.1, 0.9], [0.1, 0.9]]).to(DEVICE)

targ_0_soft = torch.tensor([[0.1, 0.9], [0.9, 0.1], [0.9, 0.1]]).to(DEVICE)

targ_mixed_soft = torch.tensor([[0.9, 0.1], [0.1, 0.9], [0.9, 0.1]]).to(DEVICE)



class_weight = torch.tensor([1.0, 2.0]).to(DEVICE)





def test_wbce():

    """ """

    criterion_wbce = WeightedBCELoss(torch.ones((1, 2)), PosWeightIsDynamic=True).to(DEVICE)

    assert criterion_wbce(torch.sigmoid(inp), targ_1).item() == pytest.approx(0.0, abs=1e-4)

    assert criterion_wbce(torch.sigmoid(inp), targ_0).item() > 1.0

    assert criterion_wbce(torch.sigmoid(inp), targ_mixed).item() > 1.0 / 3

    assert criterion_wbce(torch.sigmoid(inp), targ_1_soft).item() > criterion_wbce(torch.sigmoid(inp), targ_1).item()

    assert criterion_wbce(torch.sigmoid(inp), targ_0_soft).item() < criterion_wbce(torch.sigmoid(inp), targ_0).item()

    assert (

        criterion_wbce(torch.sigmoid(inp), targ_1).item()

        < criterion_wbce(torch.sigmoid(inp), targ_mixed_soft).item()

        < criterion_wbce(torch.sigmoid(inp), targ_0).item()

    )

    assert (

        criterion_wbce(torch.sigmoid(inp), targ_1_soft).item()

        < criterion_wbce(torch.sigmoid(inp), targ_mixed_soft).item()

        < criterion_wbce(torch.sigmoid(inp), targ_0_soft).item()

    )



    criterion_wbce = WeightedBCELoss(torch.ones((1, 2)), reduce=False).to(DEVICE)

    criterion_wbce(torch.sigmoid(inp), targ_1)

    criterion_wbce = WeightedBCELoss(torch.ones((1, 2)), size_average=False).to(DEVICE)

    criterion_wbce(torch.sigmoid(inp), targ_1)



    with pytest.raises(ValueError, match="Target size \\(.+\\) must be the same as input size \\(.+\\)"):

        criterion_wbce = WeightedBCELoss(torch.ones((1, 2))).to(DEVICE)

        criterion_wbce(torch.sigmoid(inp), targ_1[:, 0:1])





def test_bce_cw():

    """ """

    criterion_bce_cw = BCEWithLogitsWithClassWeightLoss(class_weight=class_weight).to(DEVICE)

    for targ in [targ_1, targ_0, targ_1_soft, targ_0_soft]:

        loss_1 = criterion_bce_cw(inp, targ)

        loss_2 = -class_weight * (targ * torch.log(torch.sigmoid(inp)) + (1 - targ) * torch.log(1 - torch.sigmoid(inp)))

        loss_2 = loss_2.mean()

        assert torch.allclose(loss_1, loss_2, atol=1e-3)





def test_focal():

    """ """

    criterion_focal = FocalLoss(class_weight=class_weight, multi_label=True).to(DEVICE)

    assert criterion_focal(inp, targ_1).item() == pytest.approx(0.0, abs=1e-6)

    assert criterion_focal(inp, targ_0).item() > 1.0

    assert criterion_focal(inp, targ_mixed).item() > 1.0 / 3

    assert criterion_focal(inp, targ_1_soft).item() > criterion_focal(inp, targ_1).item()

    assert criterion_focal(inp, targ_0_soft).item() < criterion_focal(inp, targ_0).item()

    assert (

        criterion_focal(inp, targ_1).item() < criterion_focal(inp, targ_mixed_soft).item() < criterion_focal(inp, targ_0).item()

    )

    assert (

        criterion_focal(inp, targ_1_soft).item()

        < criterion_focal(inp, targ_mixed_soft).item()

        < criterion_focal(inp, targ_0_soft).item()

    )

    assert torch.allclose(criterion_focal.alpha, class_weight, atol=1e-3)



    criterion_focal = FocalLoss(class_weight=class_weight.unsqueeze(0), multi_label=False, reduction="sum").to(DEVICE)

    criterion_focal(inp, targ_1)





def test_asl():

    """ """

    criterion_asl = AsymmetricLoss().to(DEVICE)

    assert criterion_asl(inp, targ_1).item() == pytest.approx(0.0, abs=1e-6)

    assert criterion_asl(inp, targ_0).item() > 1.0

    assert criterion_asl(inp, targ_mixed).item() > 1.0 / 3

    assert criterion_asl(inp, targ_1_soft).item() > criterion_asl(inp, targ_1).item()

    assert criterion_asl(inp, targ_0_soft).item() < criterion_asl(inp, targ_0).item()

    assert criterion_asl(inp, targ_1).item() < criterion_asl(inp, targ_mixed_soft).item() < criterion_asl(inp, targ_0).item()

    assert (

        criterion_asl(inp, targ_1_soft).item()

        < criterion_asl(inp, targ_mixed_soft).item()

        < criterion_asl(inp, targ_0_soft).item()

    )



    criterion_asl = AsymmetricLoss(implementation="deep-psp").to(DEVICE)

    assert criterion_asl(inp, targ_1).item() == pytest.approx(0.0, abs=1e-6)

    assert criterion_asl(inp, targ_0).item() > 1.0

    assert criterion_asl(inp, targ_mixed).item() > 1.0 / 3

    assert criterion_asl(inp, targ_1_soft).item() > criterion_asl(inp, targ_1).item()

    assert criterion_asl(inp, targ_0_soft).item() < criterion_asl(inp, targ_0).item()

    assert criterion_asl(inp, targ_1).item() < criterion_asl(inp, targ_mixed_soft).item() < criterion_asl(inp, targ_0).item()

    assert (

        criterion_asl(inp, targ_1_soft).item()

        < criterion_asl(inp, targ_mixed_soft).item()

        < criterion_asl(inp, targ_0_soft).item()

    )



    criterion_asl = AsymmetricLoss(disable_torch_grad_focal_loss=True, reduction="sum").to(DEVICE)

    criterion_asl(inp, targ_1)



    criterion_asl = AsymmetricLoss(disable_torch_grad_focal_loss=True, reduction="none").to(DEVICE)

    criterion_asl(inp, targ_1)



    with pytest.raises(ValueError, match="`prob_margin` must be non-negative"):

        AsymmetricLoss(prob_margin=-0.1)





def test_mbce():

    criterion_mbce = MaskedBCEWithLogitsLoss().to(DEVICE)

    weight_mask = torch.ones_like(inp).to(DEVICE)

    weight_mask[:, 0] = 10.0

    assert criterion_mbce(inp, targ_1, weight_mask).item() == pytest.approx(0.0, abs=1e-3)

    assert criterion_mbce(inp, targ_0, weight_mask).item() > 10.0 / 3

    assert criterion_mbce(inp, targ_mixed, weight_mask).item() > 10.0 / 3

    assert criterion_mbce(inp, targ_1_soft, weight_mask).item() > criterion_mbce(inp, targ_1, weight_mask).item()

    assert criterion_mbce(inp, targ_0_soft, weight_mask).item() < criterion_mbce(inp, targ_0, weight_mask).item()

    assert (

        criterion_mbce(inp, targ_1, weight_mask).item()

        < criterion_mbce(inp, targ_mixed_soft, weight_mask).item()

        < criterion_mbce(inp, targ_0, weight_mask).item()

    )





def test_setup_criterion():

    criterion = setup_criterion("WeightedBCELoss", pos_weight=torch.ones((1, 2)))

    criterion = setup_criterion("BCEWithLogitsWithClassWeightLoss", class_weight=class_weight)

    criterion = setup_criterion("FocalLoss", class_weight=class_weight)

    criterion = setup_criterion("AsymmetricLoss")

    criterion = setup_criterion("MaskedBCEWithLogitsLoss")



    for name in torch.nn.modules.loss.__all__:

        criterion = setup_criterion(name)