def inference(self, x_input, interpolation="nearest"):
x_img, x_pass = self.get_features(x_input)
o_height, o_width = x_pass.shape[:2]
x_height, x_width = x_img.shape[:2]
x_img = np.reshape(x_img, (x_height * x_width, -1))
probs = self.clf.predict_proba(x_img)
n_classes = probs.shape[1]
y_img = []
for i in range(n_classes):
y_i_img = np.reshape(probs[:, i], (x_height, x_width, 1))
y_img.append(y_i_img)
y_img = np.concatenate(y_img, axis=2)
x_img_pass = resize(x_pass,
width=o_width,
height=o_height,
interpolation="nearest")
y_img = resize(y_img,
width=o_width,
height=o_height,
interpolation=interpolation)
if len(x_img_pass.shape) < 3:
x_img_pass = np.expand_dims(x_img_pass, axis=2)
if len(y_img.shape) < 3:
y_img = np.expand_dims(y_img, axis=2)
y_img = np.concatenate([x_img_pass, y_img], axis=2)
return y_img
def inference(self, x_input, interpolation="nearest"):
segments = generate_segments(x_input, self.opt)
x_img, x_pass = self.get_features(x_input)
o_height, o_width = x_pass.shape[:2]
x_height, x_width = x_img.shape[:2]
x_img = get_features_for_segments(x_img, segments,
self.opt["feature_aggregation"])
y_pred = self.clf.predict_proba(x_img)
segments = resize(segments,
width=x_width,
height=x_height,
interpolation="nearest")
y_img = map_segments(segments, y_pred)
x_img_pass = resize(x_pass,
width=o_width,
height=o_height,
interpolation="nearest")
y_img = resize(y_img,
width=o_width,
height=o_height,
interpolation=interpolation)
if len(x_img_pass.shape) < 3:
x_img_pass = np.expand_dims(x_img_pass, axis=2)
if len(y_img.shape) < 3:
y_img = np.expand_dims(y_img, axis=2)
y_img = np.concatenate([x_img_pass, y_img], axis=2)
return y_img
def predict(self, tag):
x_input = tag.load_x()
segments = generate_segments(x_input, self.opt)
x_img, x_pass = self.get_features(tag)
o_height, o_width = x_pass.shape[:2]
x_height, x_width = x_img.shape[:2]
x_img = get_features_for_segments(x_img, segments, self.opt["feature_aggregation"])
y_img = self.clf.predict(x_img)
segments = resize(segments, width=x_width, height=x_height, interpolation="nearest")
y_img = map_segments(segments, y_img)
y_img = resize(y_img, width=o_width, height=o_height, interpolation="nearest")
return y_img
def transform_features(self, x_img):
x = resize(x_img,
width=self.global_kernel[0],
height=self.global_kernel[1],
interpolation="linear")
x = np.reshape(x, (1, -1))
return x
def inference(self, x_img):
if self.parameter is None:
return x_img
height, width = x_img.shape[:2]
return resize(x_img,
width=int(width / self.parameter),
height=int(height / self.parameter))
def inference(self, x_input, interpolation="nearest"):
x_img = self.get_features(x_input)
o_h, o_w = x_img.shape[:2]
x_img = resize(x_img,
height=int(o_h / 2**self.down_scale),
width=int(o_w / 2**self.down_scale))
x_img = normalize(x_img)
h, w = x_img.shape[:2]
if len(x_img.shape) < 3:
x_img = np.expand_dims(x_img, axis=2)
num_f = x_img.shape[2]
x_int_mm = np.zeros((h, w, num_f))
x_int_pp = np.zeros((h, w, num_f))
x_int_mp = np.zeros((h, w, num_f))
x_int_pm = np.zeros((h, w, num_f))
step = w + h
for i in range(1, w - 1, 1):
for j in range(1, h - 1, 1):
n = w - 1 - i
k = h - 1 - j
i_m = i - 1
j_m = j - 1
n_p = n + 1
k_p = k + 1
x_int_mm[j, i, :] = x_img[j, i, :] / step + (
x_int_mm[j_m, i_m, :] + x_int_mm[j_m, i, :] +
x_int_mm[j, i_m, :]) / 3
x_int_pp[k, n, :] = x_img[k, n, :] / step + (
x_int_pp[k_p, n_p, :] + x_int_pp[k_p, n, :] +
x_int_pp[k, n_p, :]) / 3
x_int_mp[j, n, :] = x_img[j, n, :] / step + (
x_int_mp[j_m, n_p, :] + x_int_mp[j_m, n, :] +
x_int_mp[j, n_p, :]) / 3
x_int_pm[k, i, :] = x_img[k, i, :] / step + (
x_int_pm[k_p, i_m, :] + x_int_pm[k_p, i, :] +
x_int_pm[k, i_m, :]) / 3
x_out = [x_img, x_int_mm, x_int_mp, x_int_pm, x_int_pp]
x_out = np.concatenate(x_out, axis=2)
x_out = resize(x_out, height=o_h, width=o_w, interpolation="linear")
return x_out
def get_y_for_segments(y_img, segments):
y = []
segments = resize(segments, width=y_img.shape[1], height=y_img.shape[0])
for u in np.unique(segments):
counts = np.bincount(y_img[segments == u].astype(np.int))
y_seg = np.argmax(counts)
y.append(y_seg)
return np.array(y)
def inference(self, tag_3d, interpolation="nearest"):
x_img, x_pass = self.get_features(tag_3d)
o_height, o_width = x_pass.shape[:2]
x_height, x_width = x_img.shape[:2]
x_img = self.transform_features(x_img)
shape_parameters = self.clf.predict(x_img)
y_img = self.shape_parameters_to_label_map(shape_parameters[0], x_height, x_width)
x_img_pass = resize(x_pass, width=o_width, height=o_height, interpolation="nearest")
y_img = resize(y_img, width=o_width, height=o_height, interpolation=interpolation)
if len(x_img_pass.shape) < 3:
x_img_pass = np.expand_dims(x_img_pass, axis=2)
if len(y_img.shape) < 3:
y_img = np.expand_dims(y_img, axis=2)
y_img = np.concatenate([x_img_pass, y_img], axis=2)
return y_img
def predict(self, x_input):
x_img, x_pass = self.get_features(x_input)
o_height, o_width = x_pass.shape[:2]
y_img = self.pipeline.inference(x_img)
y_img = resize(y_img,
width=o_width,
height=o_height,
interpolation="nearest")
return y_img
def predict(self, tag_3d):
x_img, x_pass = self.get_features(tag_3d)
o_height, o_width = x_pass.shape[:2]
x_height, x_width = x_img.shape[:2]
x_img = self.transform_features(x_img)
shape_parameters = self.clf.predict(x_img)
y_img = self.shape_parameters_to_label_map(shape_parameters[0], x_height, x_width)
y_img = resize(y_img, width=o_width, height=o_height, interpolation="nearest")
return y_img
def eval(self, x_img, y_img):
p_img = self.inference(x_img)
h_img, w_img = p_img.shape[:2]
y_img = resize(y_img, w_img, h_img)
y_img = np.reshape(y_img, h_img * w_img)
p_img = np.reshape(p_img, h_img * w_img)
self.stats["TP"] += np.sum(np.logical_and(y_img == 1, p_img == 1))
self.stats["FP"] += np.sum(np.logical_and(y_img != 1, p_img == 1))
self.stats["FN"] += np.sum(np.logical_and(y_img == 1, p_img != 1))
def predict(self, x_input):
x_img, x_pass = self.get_features(x_input)
o_height, o_width = x_pass.shape[:2]
x_height, x_width = x_img.shape[:2]
x_img = np.reshape(x_img, (x_height * x_width, -1))
y_img = self.clf.predict(x_img)
y_img = np.reshape(y_img, (x_height, x_width))
y_img = resize(y_img,
width=o_width,
height=o_height,
interpolation="nearest")
return y_img
def inference(self, x_input, interpolation="nearest"):
x_img, x_pass = self.get_features(x_input)
o_height, o_width = x_pass.shape[:2]
x_height, x_width = x_img.shape[:2]
y_img = self.pipeline.inference(x_img)
y_img = np.reshape(y_img, (x_height, x_width, 1))
x_img_pass = resize(x_pass,
width=o_width,
height=o_height,
interpolation="nearest")
y_img = resize(y_img,
width=o_width,
height=o_height,
interpolation=interpolation)
if len(x_img_pass.shape) < 3:
x_img_pass = np.expand_dims(x_img_pass, axis=2)
if len(y_img.shape) < 3:
y_img = np.expand_dims(y_img, axis=2)
y_img = np.concatenate([x_img_pass, y_img], axis=2)
return y_img
def get_features_for_segments(tensor, segments, feature_aggregation):
if len(tensor.shape) < 3:
tensor = np.expand_dims(tensor, axis=2)
h, w, num_f = tensor.shape
segments = resize(segments, width=w, height=h)
f_tensor = np.reshape(tensor, (w * h, num_f))
f_segments = np.reshape(segments, (w * h))
if "quantiles" == feature_aggregation:
return get_features_opt_quantiles(f_tensor, f_segments)
if "hist" in feature_aggregation:
bins = int(feature_aggregation.replace("hist", ""))
return get_features_opt_histogram(f_tensor, f_segments, bins)
if "sum" == feature_aggregation:
return get_features_opt_sum(f_tensor, f_segments)
if "gauss" == feature_aggregation:
return get_features_opt_gauss(f_tensor, f_segments)
if "hog" == feature_aggregation:
return get_features_opt_hog(tensor, segments)
raise ValueError(
"FeatureAggregation - {} - not known!".format(feature_aggregation))
def predict(self, x_input):
x_img, x_pass = self.get_features(x_input)
o_height, o_width = x_pass.shape[:2]
x_height, x_width = x_img.shape[:2]
x_img = x_img[:, :, -1]
label_img = label(x_img)
props = regionprops(label_img)
x_props = []
y_img = np.zeros((x_height, x_width))
for p in props:
x_p = self.properties_to_features(p)
x_props.append(x_p)
if len(x_props) == 0:
return y_img
x_props = np.concatenate(x_props, axis=0)
y_props = self.clf.predict(x_props)
for i, p in enumerate(props):
y_img[label_img == p["label"]] = y_props[i]
y_img = resize(y_img,
width=o_width,
height=o_height,
interpolation="nearest")
return y_img
def inference(self, x_input, interpolation="nearest"):
o_height, o_width = x_input.shape[:2]
x_img = self.get_features(x_input)
x_height, x_width = x_img.shape[:2]
x_img_row = self.format_to_axis_view(np.copy(x_img), axis=1)
x_img_row = np.reshape(x_img_row, (x_height, -1))
x_img_col = self.format_to_axis_view(np.copy(x_img), axis=0)
x_img_col = np.reshape(x_img_col, (x_width, -1))
y_row = self.clf_row.predict(x_img_row)
y_col = self.clf_col.predict(x_img_col)
y_row = np.repeat(y_row, axis=1)
y_col = np.repeat(y_col, axis=0)
n_classes = probs.shape[1]
y_img = []
for i in range(n_classes):
y_i_img = np.reshape(probs[:, i], (x_height, x_width, 1))
y_img.append(y_i_img)
y_img = np.concatenate(y_img, axis=2)
y_img = resize(y_img,
width=o_width,
height=o_height,
interpolation=interpolation)
if len(x_input.shape) < 3:
x_input = np.expand_dims(x_input, axis=2)
if len(y_img.shape) < 3:
y_img = np.expand_dims(y_img, axis=2)
y_img = np.concatenate([x_input, y_img], axis=2)
return y_img
def transform_features(self, x_img):
assert x_img.shape[2] < 512, "Too many features! - {} -".format(x_img.shape[2])
x = resize(x_img, width=self.global_kernel[0], height=self.global_kernel[1], interpolation="linear")
x = np.reshape(x, (1, -1))
return x