def test_integer_index(size: int, data: st.SearchStrategy):
index = data.draw(integer_index(size), label="index")
x = np.empty((size,))
o = x[index] # raises if invalid index
assert isinstance(
o, Real
), "An integer index should produce a number from a 1D array"
def test_nan_in_grad(
self,
inputs: ndarray,
alpha: float,
gamma: float,
dtype: torch.dtype,
data: st.SearchStrategy,
):
""" Ensures, across a wide range of inputs, that the focal loss gradient is not nan. """
targets = data.draw(
hnp.arrays(
dtype=int,
shape=(inputs.shape[0], ),
elements=st.integers(0, inputs.shape[1] - 1),
),
label="targets",
)
inputs = tensor(inputs, dtype=dtype, requires_grad=True)
targets = tensor(targets, dtype=torch.long)
loss = softmax_focal_loss(inputs, targets, alpha=alpha, gamma=gamma)
loss.backward()
assert not np.any(np.isnan(
inputs.grad.numpy())), "focal loss gradient is nan"
def test_broadcast_compat_shape(
shape: Tuple[int, ...],
allow_singleton: bool,
min_dim: int,
min_side: int,
data: st.SearchStrategy,
):
""" Ensures that the `broadcastable_shape` strategy:
- produces broadcastable shapes
- respects input parameters"""
max_side = data.draw(st.integers(min_side, min_side + 5), label="max side")
max_dim = data.draw(
st.integers(min_dim, max(min_dim, len(shape) + 3)), label="max dim"
)
compat_shape = data.draw(
broadcastable_shape(
shape=shape,
allow_singleton=allow_singleton,
min_dim=min_dim,
max_dim=max_dim,
min_side=min_side,
max_side=max_side,
),
label="broadcastable_shape",
)
assert (
min_dim <= len(compat_shape) <= max_dim
), "a shape of inappropriate dimensionality was generated by the strategy"
a = np.empty(shape)
b = np.empty(compat_shape)
np.broadcast(a, b) # error if drawn shape for b is not broadcast-compatible
if not allow_singleton:
small_dim = min(a.ndim, b.ndim)
if small_dim:
assert (
shape[-small_dim:] == compat_shape[-small_dim:]
), "singleton dimensions were included by the strategy"
if len(compat_shape) > len(shape):
n = len(compat_shape) - len(shape)
for side in compat_shape[:n]:
assert (
min_side <= side <= max_side
), "out-of-bound sides were generated by the strategy"
def test_choices(seq: List[int], replace: bool, data: st.SearchStrategy):
""" Ensures that the `choices` strategy:
- draws from the provided sequence
- respects input parameters"""
upper = len(seq) + 10 if replace and seq else len(seq)
size = data.draw(st.integers(0, upper), label="size")
chosen = data.draw(choices(seq, size=size, replace=replace), label="choices")
assert set(chosen) <= set(seq), (
"choices contains elements that do not " "belong to `seq`"
)
assert len(chosen) == size, "the number of choices does not match `size`"
if not replace and len(set(seq)) == len(seq):
unique_choices = sorted(set(chosen))
assert unique_choices == sorted(chosen), (
"`choices` with `replace=False` draws " "elements with replacement"
)
def test_basic_index(shape: Tuple[int, ...], data: st.SearchStrategy):
min_dim = data.draw(st.integers(0, len(shape) + 2), label="min_dim")
max_dim = data.draw(st.integers(min_dim, min_dim + len(shape)), label="max_dim")
index = data.draw(
basic_index(shape=shape, min_dim=min_dim, max_dim=max_dim), label="index"
)
x = np.zeros(shape, dtype=int)
o = x[index] # raises if invalid index
note("`x[index]`: {}".format(o))
if o.size and o.ndim > 0:
assert np.shares_memory(x, o), (
"The basic index should produce a " "view of the original array."
)
assert min_dim <= o.ndim <= max_dim, (
"The dimensionality input constraints " "were not obeyed"
)
def choose_metrics(self, num_train_metrics: int, num_test_metrics: int,
data: st.SearchStrategy):
assume(num_train_metrics + num_test_metrics > 0)
self.train_metric_names = ["metric-a", "metric-b",
"metric-c"][:num_train_metrics]
self.test_metric_names = ["metric-a", "metric-b",
"metric-c"][:num_test_metrics]
train_colors = data.draw(
st.lists(
cst.matplotlib_colors(),
min_size=num_train_metrics,
max_size=num_train_metrics,
),
label="train_colors",
)
test_colors = data.draw(
st.lists(
cst.matplotlib_colors(),
min_size=num_test_metrics,
max_size=num_test_metrics,
),
label="test_colors",
)
metrics = OrderedDict((n, dict()) for n in sorted(
set(self.train_metric_names + self.test_metric_names)))
for metric, color in zip(self.train_metric_names, train_colors):
metrics[metric]["train"] = color
for metric, color in zip(self.test_metric_names, test_colors):
metrics[metric]["test"] = color
self.plotter = LivePlot(
metrics,
max_fraction_spent_plotting=data.draw(
st.floats(0, 1), label="max_fraction_spent_plotting"),
last_n_batches=data.draw(st.none() | st.integers(1, 100),
label="last_n_batches"),
)
self.logger = LiveLogger()
note("Train metric names: {}".format(self.train_metric_names))
note("Test metric names: {}".format(self.test_metric_names))
def set_test_batch(self, batch_size: int, data: SearchStrategy):
self.test_batch_set = True
batch = {
name: data.draw(st.floats(-1, 1), label=name)
for name in self.test_metric_names
}
self.logger.set_test_batch(metrics=batch, batch_size=batch_size)
self.plotter.set_test_batch(metrics=batch, batch_size=batch_size)
def test_logsumexp(data: st.SearchStrategy, x: np.ndarray, keepdims: bool):
axes = data.draw(valid_axes(ndim=x.ndim), label="axes")
mygrad_result = logsumexp(x, axis=axes, keepdims=keepdims)
scipy_result = special.logsumexp(x, axis=axes, keepdims=keepdims)
assert_array_equal(
mygrad_result,
scipy_result,
err_msg="mygrad's implementation of logsumexp does "
"not match that of scipy's",
)
def test_comparison_ops(
op: str, x: np.ndarray, x_constant: bool, y_constant: bool, data: st.SearchStrategy
):
y = data.draw(hnp.arrays(shape=x.shape, dtype=x.dtype, elements=st.floats(-10, 10)))
x = Tensor(x, constant=x_constant)
y = Tensor(y, constant=y_constant)
assert hasattr(Tensor, op), "`Tensor` is missing the attribute {}".format(op)
tensor_out = getattr(Tensor, op)(x, y)
array_out = getattr(np.ndarray, op)(x.data, y.data)
assert_equal(actual=tensor_out, desired=array_out)
def set_test_batch(self, batch_size: int, data: SearchStrategy):
self.test_batch_set = True
batch = {
metric.name: data.draw(
st.floats(-1, 1) | st.floats(-1, 1).map(np.array), label=metric.name
)
for metric in self.test_metrics
}
self.logger.set_test_batch(metrics=batch, batch_size=batch_size)
for metric in self.test_metrics:
metric.add_datapoint(batch[metric.name], weighting=batch_size)
def test_matches_simple_implementation(
self,
inputs: ndarray,
alpha: float,
gamma: float,
dtype: torch.dtype,
data: st.SearchStrategy,
):
""" Ensures that focal loss matches a naive-implementation over the domain where numerical
stability is not an issue. """
targets = data.draw(
hnp.arrays(
dtype=int,
shape=(inputs.shape[0], ),
elements=st.integers(0, inputs.shape[1] - 1),
),
label="targets",
)
inputs1 = tensor(inputs, dtype=dtype, requires_grad=True)
inputs2 = tensor(inputs, dtype=dtype, requires_grad=True)
targets = tensor(targets, dtype=torch.int64)
# numerically-stable focal loss
loss = softmax_focal_loss(inputs1, targets, alpha=alpha, gamma=gamma)
loss.backward()
# naive focal loss
input = F.softmax(inputs2, dim=1)
pc = input[(range(len(targets)), targets)]
naive_loss = (-alpha * (1 - pc)**gamma * torch.log(pc)).mean()
naive_loss.backward()
assert_allclose(
actual=loss.detach().numpy(),
desired=naive_loss.detach().numpy(),
atol=1e-5,
rtol=1e-5,
err_msg="focal loss does not match naive implementation on "
"numerically-stable domain",
)
assert_allclose(
actual=inputs1.grad.numpy(),
desired=inputs2.grad.numpy(),
atol=1e-5,
rtol=1e-5,
err_msg="focal loss gradient does not match that of naive loss on "
"numerically-stable domain",
)
def test_slice_index(size: int, data: st.SearchStrategy):
index = data.draw(slice_index(size), label="index")
x = np.empty((size,))
o = x[index] # raises if invalid index
assert isinstance(o, np.ndarray) and o.ndim == 1, (
"A slice index should produce " "a 1D array from a 1D array"
)
if o.size:
assert np.shares_memory(o, x), "A slice should produce a view of `x`"
if index.start is not None:
assert -size <= index.start <= size
if index.stop is not None:
assert -size <= index.stop <= size
def test_shapes(self, boxes: ndarray, truth: ndarray,
data: st.SearchStrategy):
""" Ensure the shape returned by generate_targets is correct, even in edge cases producing empty arrays. """
boxes = boxes.cumsum(
axis=1) # to ensure we don't hit 0-width or -height boxes
truth = truth.cumsum(
axis=1) # to ensure we don't hit 0-width or -height boxes
N = boxes.shape[0]
K = truth.shape[0]
labels = data.draw(hnp.arrays(dtype=int, shape=(K, )))
cls, reg = generate_targets(boxes, truth, labels, 0.5, 0.4)
msg = "generate_targets failed to produce classification targets of the correct shape"
assert cls.shape == (N, ), msg
msg = "generate_targets failed to produce regression targets of the correct shape"
assert reg.shape == (N, 4), msg
def test_matches_crossentropy(self, inputs: ndarray, alpha: float,
dtype: torch.dtype, data: st.SearchStrategy):
""" Ensures that focal loss w/ gamma=0 matches softmax cross-entropy (scaled by alpha). """
targets = data.draw(
hnp.arrays(
dtype=int,
shape=(inputs.shape[0], ),
elements=st.integers(0, inputs.shape[1] - 1),
),
label="targets",
)
inputs = tensor(inputs, dtype=dtype)
targets = tensor(targets, dtype=torch.long)
assert_allclose(
desired=alpha * F.cross_entropy(inputs, targets),
actual=softmax_focal_loss(inputs, targets, alpha=alpha, gamma=0.0),
atol=1e-6,
rtol=1e-6,
err_msg=
"Focal loss with gamma=0 fails to match cross-entropy loss.",
)
