def predict(net, filename):
'''Run neural net on images from image_set'''
file_name = PREFIX_IMAGES + filename + '.png'
pred = nn.classify(file_name, net, NN_PROB_THRESH)
car_preds = []
for p in pred:
if p[0] == 0:
car_preds.append(
[0, p[1]] +
[int(p[2][0]),
int(p[2][1]),
int(p[2][2]),
int(p[2][3])])
# car_preds.append([int(p[2][0]), int(p[2][1]), int(p[2][2]), int(p[2][3])])
for p in car_preds:
scale_x = 1.5385
scale_y = 3.125
p[2] = int(p[2] * scale_x)
p[3] = int(p[3] * scale_y)
p[4] = int(p[4] * scale_x)
p[5] = int(p[5] * scale_y)
# print("car_preds: ")
# print(car_preds)
return car_preds
def predict(net, image_set):
'''Run neural net on images from image_set'''
with open(image_set, "r") as f:
content = f.readlines()
images = [c.strip() for c in content]
predictions = dict()
for i in images:
file_name = PREFIX_IMAGES + i + '.png'
pred = nn.classify(file_name, net, NN_PROB_THRESH)
boxes = []
for p in pred:
if p[0] == 0:
boxes += [list(p[2])]
for box in boxes:
# Matrix
#scale_x = 1.5336
#scale_y = 2.7396
# Scenic
scale_x = 1.5385
scale_y = 3.125
box[0] = int(box[0] * scale_x)
box[1] = int(box[1] * scale_y)
box[2] = int(box[2] * scale_x)
box[3] = int(box[3] * scale_y)
predictions[i] = boxes
return predictions
def predict(net, image_set):
'''Run neural net on images from image_set'''
with open(image_set, "r") as f:
content = f.readlines()
images = [c.strip() for c in content]
predictions = dict()
for i in images:
file_name = PREFIX_IMAGES + i + '.jpg'
pred = nn.classify(file_name, net, NN_PROB_THRESH)
car_preds = []
for p in pred:
if p[0] == 0:
car_preds.append([0, p[1]] + [int(p[2][0]), int(p[2][1]), int(p[2][2]), int(p[2][3])])
for p in car_preds:
# Matrix
scale_x = 1.5336
scale_y = 2.7396
# Scenic
#scale_x = 1.5385
#scale_y = 3.125
p[2] = int(p[2]*scale_x)
p[3] = int(p[3]*scale_y)
p[4] = int(p[4]*scale_x)
p[5] = int(p[5]*scale_y)
predictions[i] = car_preds
return predictions
def predict(net, image_set):
'''Run neural net on images from image_set'''
with open(image_set, "r") as f:
content = f.readlines()
images = [c.strip() for c in content]
predictions = dict()
for i in images:
(boxes, probs, _) = nn.classify(PREFIX_IMAGES + i + '.png', net,
NN_PROB_THRESH)
predictions[i] = boxes
return predictions
#CHECKPOINT = '/home/tommaso/scenicEx/data/scenic/checkpoints/train/model.ckpt-5500' #CHECKPOINT = '/home/tommaso/scenicEx/data/matrix/checkpoints_overlap_250/train/model.ckpt-32670' #CHECKPOINT = '/home/tommaso/scenicEx/data/matrix/checkpoints/train/model.ckpt-5250' CHECKPOINT = '/home/tommaso/squeezeDet/data/webots/train/checkpoints/train/model.ckpt-1150' FILE = '/home/tommaso/squeezeDet/data/webots/test/images/1003.png' PREFIX = '/home/tommaso/scenicEx/data/matrix/' PREFIX_LABELS = PREFIX + 'labels/' PREFIX_IMAGES = PREFIX + 'images/' image_set = '/home/tommaso/scenicEx/data/matrix/ImageSets/test_overlap.txt' net = nn.init(CHECKPOINT) pred = nn.classify(FILE, net, NN_PROB_THRESH) print(pred) #ap, recall = eval_set(net, image_set) # avg_precs = [] # avg_recs = [] # max_precs = [] # max_recs = [] # # for i in range(1,9,1): # precs = [] # recs = [] # for j in range(4950,5020,10): # checkpoint_path = '/home/tommaso/scenicEx/data/matrix/checkpoints_overlap_250/checkpoints_overlap_250_' + str(i) + '/train/model.ckpt-' + str(j) # print(checkpoint_path) # net = nn.init(checkpoint_path)
'''Compute precision and recall'''
tp = sum(d != -1 for d in detects)
fp = sum(d == -1 for d in detects)
fn = sum(g not in detects for g in range(n_gt))
print((tp, fp, fn))
prec = tp / float(tp + fp)
rec = tp / float(tp + fn)
return prec, rec
def precision_recall(preds, gt_boxes, iou_tresh):
detects = pred_2_detect(preds, gt_boxes, iou_tresh)
return prec_rec(detects, len(gt_boxes))
checkpoint = '/home/tommaso/analyzeNN/data/train_0/checkpoint/train/model.ckpt-5000'
net = nn.init(checkpoint)
tmp_img_file_name = "/home/tommaso/analyzeNN/data/test/images/h_m_0_3_000006.png"
gt_b_1 = edge_2_cent([641, 181, 675, 209])
gt_b_2 = edge_2_cent([598, 178, 644, 214])
gt_b_3 = edge_2_cent([514, 162, 604, 236])
gt_boxes = [gt_b_1, gt_b_2, gt_b_3]
(boxes, probs, cats) = nn.classify(tmp_img_file_name, net)
def main(argv):
print(argv)
TEST_NN = True
reaction_time = 0.75 if TEST_NN else float(argv[1])
cruising_speed = 15.0 if TEST_NN else float(argv[2])
print(reaction_time)
steering_constant = 35.0 * math.pow(cruising_speed, -1.442)
mean = 0.55
var = 0.5
# First the car does lane keeping for about 1.5 secs (50 iters)
driver.setCruisingSpeed(cruising_speed)
driver.setSteeringAngle(0)
driver.step()
# * np.sqrt(2 * np.pi * var)
x_start = mean + np.sqrt(2 * var) * erfinv(-0.99)
x_end = mean + np.sqrt(2 * var) * erfinv(0.99)
max_iters = int(math.ceil((x_end - x_start) / delta_t))
reaction_iters = int(math.ceil(reaction_time / delta_t))
print("Mean, var, reaction time: ", mean, var, reaction_time)
d = np.inf
iter = 0
k = 2
offset_center = 0
print("Mode: LANE KEEPING")
#while d > k*cruising_speed:
while d > 15.:
#while True:
if iter % DETECT_FREQ == 1:
frame = get_camera_image(camera_lk, IMG_W_LK, IMG_H_LK)
frame = np.array(frame.convert('RGB'))
try:
offset_center = process_pipeline(frame, keep_state=False)
except:
print('Lane assistant: Skipping frame')
offset_center = 0
steer_angle = offset_center * 0.5
driver.setSteeringAngle(steer_angle)
image_cones = get_camera_image(camera_cnn, IMG_W_CNN, IMG_H_CNN)
# convert to a openCV2 image
cv2_image = cv2.cvtColor(np.array(image_cones), cv2.COLOR_RGB2BGR)
# call neural net
pred = nn.classify(cv2_image, net, NN_PROB_THRESH)
d = estimate_dist(pred)
print("CNN estimated distance from cones: ", d)
iter += 1
driver.step()
# Now start the gaussian maneuver
x_current = x_start
for _ in range(reaction_iters):
driver.setCruisingSpeed(cruising_speed)
driver.setSteeringAngle(0)
driver.step()
driver.setCruisingSpeed(cruising_speed)
driver.setSteeringAngle(normal_steering(x=x_current, mean=mean, var=var))
print("Mode: LANE CHANGE")
for iter in range(max_iters + 1):
driver.setCruisingSpeed(cruising_speed)
if iter % 8 == 0:
x_current = x_start + iter * delta_t
print(
x_current,
steering_constant *
normal_steering(x=x_current, mean=mean, var=var),
x_current - mean)
driver.setSteeringAngle(
-steering_constant *
normal_steering(x=x_current, mean=mean, var=var))
driver.step()
# Now do lane keeping again
print("Mode: LANE KEEPING")
iter = 0
while driver.step() != -1:
if iter % 8 == 0:
frame = get_camera_image(camera_lk, IMG_W_LK, IMG_H_LK)
frame = np.array(frame.convert('RGB'))
try:
offset_center = process_pipeline(frame, keep_state=False)
except:
print('Lane assistant: Skipping frame')
offset_center = 0
steer_angle = offset_center * 0.5
driver.setSteeringAngle(steer_angle)
iter += 1