#% Load model
model = load_model(filepath + 'unet_exp_' + str(exp) + '.h5', compile=False)
area = 11
# Prediction
ref_final, pre_final, prob_recontructed, ref_reconstructed, mask_no_considered_, mask_ts, time_ts = prediction(
model, image_array, image_ref, final_mask, mask_ts_, patch_size, area)
# Metrics
cm = confusion_matrix(ref_final, pre_final)
metrics = compute_metrics(ref_final, pre_final)
print('Confusion matrix \n', cm)
print('Accuracy: ', metrics[0])
print('F1score: ', metrics[1])
print('Recall: ', metrics[2])
print('Precision: ', metrics[3])
# Alarm area
total = (cm[1, 1] + cm[0, 1]) / len(ref_final) * 100
print('Area to be analyzed', total)
print('training time', end_training)
print('test time', time_ts)
#%% Show the results
# prediction of the whole image
fig1 = plt.figure('whole prediction')
plt.imshow(prob_recontructed)
# Show the test tiles
fig2 = plt.figure('prediction of test set')
plt.imshow(prob_recontructed * mask_ts)
# ref_final, pre_final, prob_recontructed, ref_reconstructed, mask_no_considered_, mask_ts, time_ts = prediction(model, image_array, image_ref, final_mask, mask_ts_, patch_size, area)
# Metrics
true_labels = np.reshape(patches_test_ref, (patches_test_ref.shape[0]* patches_test_ref.shape[1]*patches_test_ref.shape[2]))
predicted_labels = np.reshape(patches_pred, (patches_pred.shape[0]* patches_pred.shape[1]*patches_pred.shape[2]))
cm = confusion_matrix(true_labels, predicted_labels)
metrics = compute_metrics(true_labels, predicted_labels)
print('Confusion matrix \n', cm)
print('Accuracy: ', metrics[0])
print('F1score: ', metrics[1])
print('Recall: ', metrics[2])
print('Precision: ', metrics[3])
# Alarm area
total = (cm[1,1]+cm[0,1])/len(true_label)*100
print('Area to be analyzed',total)
print('training time', end_training)
print('test time', time_ts)
#%% Show the results
# prediction of the whole image
fig1 = plt.figure('whole prediction')
plt.imshow(prob_recontructed)
# Show the test tiles
# fig2 = plt.figure('prediction of test set')
# plt.imshow(prob_recontructed*mask_ts)
idx = 418
for j in range(50):
idx = int(reach[j])
trainx = tss_trainx[idx + i * 100:idx + 1 + i * 100]
trainy = tss_trainy[idx + i * 100:idx + 1 + i * 100]
if not trainy:
trainy = -1
############ We use the DeepExplain repository (https://github.com/marcoancona/DeepExplain) to interpret the CNN model predictions##########
with DeepExplain(session=sess) as de:
# We run `explain()` several time to compare different attribution methods
attributions = {
# Gradient-based
'Saliency maps': de.explain('saliency', x1 * trainy, x, trainx),
#'Gradient * Input': de.explain('grad*input', x1 * trainy, x, trainx),
'Integrated Gradients': de.explain('intgrad', x1 * trainy, x,
trainx),
#'Epsilon-LRP': de.explain('elrp', logits * yi, X, xi),
#'DeepLIFT (Rescale)': de.explain('deeplift', logits * yi, X, xi),
#Perturbation-based
#'_Occlusion [1x1]': de.explain('occlusion', logits * yi, X, xi),
#'_Occlusion [3x3]': de.explain('occlusion', logits * yi, X, xi, window_shape=(3,))
}
print('Done', i)
grads.append(attributions['Integrated Gradients'][0])
grads = np.stack(grads)
grads_sum = np.mean(grads, axis=0)
im = filters.gaussian_filter(grads_sum, 4)
pdb.set_trace()
plt.imshow(im)
plt.show()
g = img_train_ref[i][j][1]
b = img_train_ref[i][j][2]
rgb = (r, g, b)
rgb_key = str(rgb)
binary_img_train_ref[i][j] = label_dict[rgb_key]
return binary_img_train_ref
root_path = './DATASETS/homework3'
# Load images
img_train_path = 'Image_Train.tif'
img_train = load_tiff_image(os.path.join(root_path,
img_train_path))
img_train = img_train.transpose((1, 2, 0))
plt.imshow(img_train)
plt.show()
print(f'Img train max: {img_train.max()}, Img train min: {img_train.min()}')
# img_train = plt.imread(os.path.join(root_path,
# img_train_path))
# print(type(img_train))
# Normalizes the image
# scaler = StandardScaler()
# scaler = MinMaxScaler(feature_range=(-1,1))
# img_reshaped = img_train.reshape((img_train.shape[0]*img_train.shape[1]), img_train.shape[2])
# scaler.fit(img_reshaped)
# image_normalized = scaler.transform(img_reshaped)
# img_normalized = image_normalized.reshape(img_train.shape[0], img_train.shape[1], img_train.shape[2])
img_normalized = img_train / 127.5 - 1
print(f'Img norm max: {img_normalized.max()}, Img norm min: {img_normalized.min()}')
# img_normalized = img_normalized.transpose((1, 2, 0))
# -*- coding: utf-8 -*-
"""
Created on Mon Jul 24 20:37:24 2017
@author: ZSHUJON
"""
from utils import imread, plt
road_image = imread('test_images/solidYellowCurve2.jpg')
plt.imshow(road_image)
(60, 60), font, 1, (255, 255, 255), 2)
cv2.putText(blend_on_road,
'Offset from center: {:.02f}m'.format(offset_meter), (60, 90),
font, 1, (255, 255, 255), 2)
processed_frames += 1
return blend_on_road
if __name__ == '__main__':
# step 1: calibrate the camera
mtx, dist = calibrate('camera_cal')
mode = "video"
processed_frames = 0
line_lt, line_rt = Line(buffer_len=10), Line(buffer_len=10)
if mode == "image":
for test_img in glob.glob('test_images/*.jpg'):
img = mpimg.imread(test_img)
img_out = process_pipeline(img, True)
plt.imshow(img_out)
plt.show()
elif mode == "video":
selector = 'project'
clip = VideoFileClip(
'{}_video.mp4'.format(selector)).fl_image(process_pipeline)
clip.write_videofile('out_{}.mp4'.format(selector), audio=False)