def test_focal_loss_fn():
# classification test
torch_ce = F.binary_cross_entropy_with_logits(INP_BINARY,
TARGET_BINARY.float())
my_ce = pt_F.focal_loss_with_logits(INP_BINARY,
TARGET_BINARY,
alpha=-1,
gamma=0)
assert torch.allclose(torch_ce, my_ce)
# check that smooth combination works
my_ce_not_reduced = pt_F.focal_loss_with_logits(INP_BINARY,
TARGET_BINARY,
combine_thr=0)
my_ce_reduced = pt_F.focal_loss_with_logits(INP_BINARY,
TARGET_BINARY,
combine_thr=0.2)
my_ce_reduced2 = pt_F.focal_loss_with_logits(INP_BINARY,
TARGET_BINARY,
combine_thr=0.8)
assert my_ce_not_reduced < my_ce_reduced
assert my_ce_reduced < my_ce_reduced2
# images test
torch_ce = F.binary_cross_entropy_with_logits(INP_IMG_BINARY,
TARGET_IMG_BINARY)
my_ce = pt_F.focal_loss_with_logits(INP_IMG_BINARY,
TARGET_IMG_BINARY,
alpha=-1,
gamma=0)
assert torch.allclose(torch_ce, my_ce)
def test_focal_loss_fn_normalize():
# simulate very accurate predictions
inp = TARGET_BINARY * 5 - (1 - TARGET_BINARY) * 5
my_ce = pt_F.focal_loss_with_logits(inp, TARGET_BINARY, normalized=False)
my_ce_normalized = pt_F.focal_loss_with_logits(inp,
TARGET_BINARY,
normalized=True)
assert my_ce_normalized > my_ce
def test_focal_loss_fn_basic():
input_good = torch.Tensor([10, -10, 10]).float()
input_bad = torch.Tensor([-1, 2, 0]).float()
target = torch.Tensor([1, 0, 1])
loss_good = pt_F.focal_loss_with_logits(input_good, target)
loss_bad = pt_F.focal_loss_with_logits(input_bad, target)
assert loss_good < loss_bad
def test_focal_loss_fn_basic():
"""explicit tests for values in two corner cases"""
input_good = torch.Tensor([10, -10, 10]).float()
input_bad = torch.Tensor([-1, 2, 0]).float()
target = torch.Tensor([1, 0, 1])
loss_good = pt_F.focal_loss_with_logits(input_good, target)
loss_bad = pt_F.focal_loss_with_logits(input_bad, target)
assert torch.allclose(loss_good, torch.tensor(0.0))
assert torch.allclose(loss_bad, torch.tensor(0.4854), atol=1e-4)
def test_focal_fn_is_scribtable():
"""check that script gives the same results"""
input_bad = torch.Tensor([-1, 2, 0]).float()
target = torch.Tensor([1, 0, 1])
loss = pt_F.focal_loss_with_logits(input_bad, target)
jit_func = torch.jit.script(pt_F.focal_loss_with_logits)
loss_jit = jit_func(input_bad, target)
assert torch.allclose(loss, loss_jit)
def test_focal_loss_fn_reduction(reduction):
torch_ce = F.binary_cross_entropy_with_logits(INP_BINARY,
TARGET_BINARY.float(),
reduction=reduction)
my_ce = pt_F.focal_loss_with_logits(INP_BINARY,
TARGET_BINARY,
alpha=0.5,
gamma=0,
reduction=reduction)
assert torch.allclose(torch_ce, my_ce * 2)
