def test_angular_weighted_illegal(self):
"""Test that only possible to use weighing when single sample."""
y_true = np.zeros((2, 10, 11, 2))
y_pred = y_true
with pytest.raises(ValueError):
angular_error_of(y_true, y_pred, weighted=True)
def test_angular_zero(self):
"""Test that zero vectors lead to nans."""
y_true = np.zeros((5, 10, 11, 2))
y_pred = np.ones_like(y_true)
assert np.all(np.isnan(angular_error_of(y_true, y_pred)[1]))
assert np.all(np.isnan(angular_error_of(y_pred, y_true)[1]))
def test_angular_weighted_vs_unweighted(self):
# One 90 (weight 2) and 180 (weight 1) degrees from the x axis
# nonweighted_average = 135, weighted = 120
delta_x_true = np.array([[0, 0], [0, -1]])
delta_y_true = np.array([[2, 0], [0, 0]])
delta_x_pred = np.ones_like(delta_x_true) # x axis is the vectors
delta_y_pred = np.zeros_like(delta_y_true)
df_true = DisplacementField(delta_x_true, delta_y_true)
df_pred = DisplacementField(delta_x_pred, delta_y_pred)
ae_nonweighted, _ = angular_error_of([df_true], [df_pred],
weighted=False)
ae_weighted, _ = angular_error_of([df_true], [df_pred], weighted=True)
assert ae_nonweighted == pytest.approx(135)
assert ae_weighted == pytest.approx(120)
def test_angular_norm_independence(self, c):
"""Test that only angle matters for `angular_error_of`."""
y_true = np.random.random((5, 10, 11, 2))
y_pred = y_true * c
metric_average, metric_per_output = angular_error_of(y_true, y_pred)
assert metric_average == pytest.approx(0, abs=1e-5)
assert np.allclose(metric_per_output, np.zeros((10, 11)), atol=1e-5)
def test_angular_norm_perpendicular(self, c):
"""Test that only angle matters for `angular_error_of`."""
y_true = np.random.random((5, 10, 11, 2))
y_pred = np.empty_like(y_true)
y_pred[..., 0] = y_true[..., 1] * c
y_pred[..., 1] = -y_true[..., 0] * c
metric_average, metric_per_output = angular_error_of(y_true, y_pred)
assert metric_average == pytest.approx(90, abs=1e-5)
assert np.allclose(metric_per_output,
90 * np.ones((10, 11)),
atol=1e-5)
def generate_df_plots(df_true, df_pred, filepath=None, figsize=(15, 15)):
"""Generate displacement vector plots.
df_true : DisplacementField
Truth. Assumes that shape is (320, 456).
df_pred : DisplacementField
Prediction. Assumes that shape is (320, 456)
filepath : None or pathlib.Path
If specified, then the path to where the figure saved as a PNG image.
If not specified, then shown.
"""
plt.ioff()
fig, (
(ax_norm, ax_norm_p),
(ax_angle, ax_angle_p),
(ax_jacob, ax_jacob_p),
(ax_grid, ax_grid_p),
) = plt.subplots(4, 2, figsize=figsize)
df_base = DisplacementField.generate(
(320, 456), approach="affine_simple",
translation_x=1) # make the angle work
bar_norm = fig.add_axes([0.95, 0.772, 0.03, 0.2])
bar_angle = fig.add_axes([0.95, 0.525, 0.03, 0.2])
bar_jacob = fig.add_axes([0.95, 0.275, 0.03, 0.2])
# Jacobian
jacob_true = df_true.jacobian
jacob_pred = df_pred.jacobian
jacob_vmin, jacob_vmax = min(jacob_true.min(), jacob_pred.min()), max(
jacob_true.max(), jacob_pred.max())
ax_jacob.set_axis_off()
sns.heatmap(
jacob_true,
ax=ax_jacob,
cbar_ax=bar_jacob,
center=0,
cmap="seismic_r",
vmin=jacob_vmin,
vmax=jacob_vmax,
)
ax_jacob.set_title("Jacobian - Ground Truth")
ax_jacob_p.set_axis_off()
sns.heatmap(
jacob_pred,
ax=ax_jacob_p,
cbar_ax=bar_jacob,
center=0,
cmap="seismic_r",
vmin=jacob_vmin,
vmax=jacob_vmax,
)
ax_jacob_p.set_title("Jacobian - Predicted")
# GRID
img_grid = create_grid((320, 456))
img_grid_warped = df_true.warp(img_grid)
ax_grid.set_axis_off()
ax_grid.imshow(img_grid_warped, cmap="gray")
ax_grid.set_title("Warped Grid - Ground Truth")
img_grid_warped_p = df_pred.warp(img_grid)
ax_grid_p.set_axis_off()
ax_grid_p.imshow(img_grid_warped_p, cmap="gray")
ax_grid_p.set_title("Warped Grid - Predicted")
# NORM
norm_vmin, norm_vmax = 0, max(df_true.norm.max(), df_pred.norm.max())
ax_norm.set_axis_off()
sns.heatmap(df_true.norm,
ax=ax_norm,
cbar_ax=bar_norm,
vmin=norm_vmin,
vmax=norm_vmax)
ax_norm.set_title("Norm - Ground Truth")
ax_norm_p.set_axis_off()
sns.heatmap(df_pred.norm,
ax=ax_norm_p,
cbar_ax=bar_norm,
vmin=norm_vmin,
vmax=norm_vmax)
ax_norm_p.set_title("Norm - Predicted")
# Angle
angle_vmin, angle_vmax = 0, 180
ax_angle.set_title("Angle wrt positive x-axis - Ground Truth")
ax_angle.set_axis_off()
angles = angular_error_of([df_true], [df_base])[1]
sns.heatmap(
angles,
ax=ax_angle,
cbar_ax=bar_angle,
mask=~np.isfinite(angles),
cmap="hot_r",
vmin=angle_vmin,
vmax=angle_vmax,
)
ax_angle_p.set_title("Angle wrt positive x-axis - Predicted")
ax_angle_p.set_axis_off()
angles = angular_error_of([df_pred], [df_base])[1]
sns.heatmap(
angles,
ax=ax_angle_p,
cbar_ax=bar_angle,
mask=~np.isfinite(angles),
cmap="hot_r",
vmin=angle_vmin,
vmax=angle_vmax,
)
# fig.tight_layout(rect=[0, 0, .95, 1])
if filepath is not None:
fig.savefig(str(filepath))
else:
plt.show()