def _prepare_data(self, X, y):
X_data = []
n = len(X[0])
m = len(X)
for i in tqdm(range(n), desc="Loading predictions"):
x_preds = []
for j in range(m):
x = np.load(X[j][i])
if self.apply_softmax == "pre":
x = torch.from_numpy(x).float().softmax(
dim=0).numpy().astype(np.float16)
x_preds.append(x)
x = np.mean(np.stack(x_preds), axis=0)
if self.apply_softmax == "post":
x = torch.from_numpy(x).float().softmax(dim=0).numpy().astype(
np.float16)
X_data.append(x)
Y_data = [
read_mask(yi) for yi in tqdm(y, desc="Loading ground-truths")
]
assert len(X_data) == len(Y_data)
print("Loaded data into memory")
return X_data, Y_data
def _prepare_data(self, X, y):
assert self.apply_softmax == "pre"
X_data = [
to_numpy(torch.from_numpy(
np.load(xi)).float().softmax(dim=0)).astype(np.float16)
for xi in X
]
Y_data = [read_mask(yi) for yi in y]
print("Loaded data into memory")
return X_data, Y_data
def test_mask_resize():
fname = "d:\\datasets\\xview2\\train\\masks\\hurricane-harvey_00000402_post_disaster.png"
mask = read_mask(fname)
mask2 = cv2.resize(mask, (512, 512), interpolation=cv2.INTER_NEAREST)
mask3 = resize_mask_one_hot(mask, (512, 512))
cv2.imshow("Original", mask * 255)
cv2.imshow("Nearest", mask2 * 255)
cv2.imshow("Smart", mask3 * 255)
cv2.waitKey(-1)
def test_watershed_with_image():
dmg = read_mask("test_damage_00121_prediction.png")
loc = read_mask("test_localization_00121_prediction.png")
img = cv2.imread("test_post_00121.png")
# Fix mask
dmg[loc == 0] = 0
seed = dmg.copy()
seed[loc == 0] = 0
markers = cv2.watershed(img, seed.astype(int))
markers[markers == 0] = 1
plt.figure()
plt.imshow(dmg)
plt.show()
plt.figure()
plt.imshow(loc)
plt.show()
plt.figure()
plt.imshow(markers)
plt.show()
def _compute_fn(args, postprocessing_fn):
xi, yi = args
dmg_pred = np.load(xi)
dmg_true = read_mask(yi)
loc_pred, dmg_pred = postprocessing_fn(dmg_pred)
if loc_pred.shape[0] != 1024:
loc_pred = cv2.resize(loc_pred,
dsize=(1024, 1024),
interpolation=cv2.INTER_NEAREST)
dmg_pred = cv2.resize(dmg_pred,
dsize=(1024, 1024),
interpolation=cv2.INTER_NEAREST)
row = CompetitionMetricCallback.get_row_pair(loc_pred, dmg_pred, dmg_true,
dmg_true)
return row
def test_watershed(actual, expected):
dmg = np.load(actual)
dmg_true = read_mask(expected)
loc_cls, dmg_cls = make_predictions_dominant_v2(dmg)
plt.figure()
plt.imshow(make_rgb_image(dmg_true))
plt.show()
plt.figure()
plt.imshow(make_rgb_image(np.argmax(dmg, axis=0)))
plt.show()
plt.figure()
plt.imshow(make_rgb_image(loc_cls))
plt.show()
plt.figure()
plt.imshow(make_rgb_image(dmg_cls))
plt.show()