def validation_score(model_path, type, save=False, debugging=False):
model = i3d(input_shape=(configs.LENGTH, configs.IMG_HEIGHT,
configs.IMG_WIDTH, configs.CHANNELS),
weights_path=model_path)
# read the steering labels and image path
df_truth = pd.read_csv(
'/home/neil/dataset/speedchallenge/data/validation.csv').values
e_sum = 0
inputs = []
predictions = []
if type == 'rgb':
start_time = time.time()
for i in range(configs.LENGTH):
file = configs.TRAIN_DIR + str(df_truth[i][1])
img = helper.load_image(file)
inputs.append(img)
# Run through all images
for i in range(configs.LENGTH, len(df_truth)):
p = configs.TRAIN_DIR + str(df_truth[i][1])
img = helper.load_image(p)
inputs.pop(0)
inputs.append(img)
prediction = model.model.predict(np.array([np.asarray(inputs)
]))[0][0]
actual_speed = df_truth[i][2]
e_sum += (actual_speed - prediction)**2
predictions.append(prediction)
if len(predictions) % 1000 == 0:
print('.')
elif type == 'rgb-flow':
print('Started')
start_time = time.time()
previous = helper.load_image(configs.TRAIN_DIR + str(df_truth[0][1]))
for i in range(1, configs.LENGTH + 1):
img = helper.load_image(configs.TRAIN_DIR + str(df_truth[i][1]))
rgbImg = helper.optical_flow_rgb(previous=previous, current=img)
if debugging:
fig = plt.figure(figsize=(3, 1))
fig.add_subplot(1, 3, 1)
plt.imshow(rgbImg)
fig.add_subplot(1, 3, 2)
plt.imshow(previous)
fig.add_subplot(1, 3, 3)
plt.imshow(img)
plt.show()
previous = img
inputs.append(rgbImg)
previous = helper.load_image(configs.TRAIN_DIR +
str(df_truth[configs.LENGTH + 1][1]))
for i in range(configs.LENGTH + 2, len(df_truth)):
img = helper.load_image(configs.TRAIN_DIR + str(df_truth[i][1]))
rgb_flow = helper.optical_flow_rgb(previous, img)
if debugging:
plt.imshow(rgb_flow)
plt.show()
inputs.pop(0)
inputs.append(rgb_flow)
input_array = np.array([np.asarray(inputs)])
prediction = model.model.predict(input_array)[0][0]
actual_steers = df_truth[i][2]
e = (actual_steers - prediction)**2
e_sum += e
predictions.append(prediction)
if len(predictions) % 1000 == 0:
print('.')
elif type == 'flow':
print('Started')
start_time = time.time()
previous = helper.load_image(configs.TRAIN_DIR + str(df_truth[0][1]))
for i in range(1, configs.LENGTH + 1):
img = helper.load_image(configs.TRAIN_DIR + str(df_truth[i][1]))
flow = helper.optical_flow(previous=previous, current=img)
inputs.append(flow)
previous = helper.load_image(configs.TRAIN_DIR +
str(df_truth[configs.LENGTH + 1][1]))
for i in range(configs.LENGTH + 2, len(df_truth)):
img = helper.load_image(configs.TRAIN_DIR + str(df_truth[i][1]))
flow = helper.optical_flow(previous, img)
inputs.pop(0)
inputs.append(flow)
prediction = model.model.predict(np.array([np.asarray(inputs)
]))[0][0]
actual_steers = df_truth[i][2]
e_sum += (actual_steers - prediction)**2
predictions.append(prediction)
if len(predictions) % 1000 == 0:
print('.')
else:
raise Exception('Sorry, the model type is not recognized')
print("time per step: %s seconds" %
((time.time() - start_time) / len(predictions)))
if save:
print("Writing predictions...")
pd.DataFrame({
"steering_angle": predictions
}).to_csv('./result.csv', index=False, header=True)
print("Done!")
return e_sum / len(predictions)
def test_loop(model_path, model_type):
"""
for visualizing acceleration models with the comma AI
validation dataset. The speed is initialized before
acceleration prediction starts.
:param model_path: the path of the trained Keras model
:param model_type: the type of model, rgb, flow or rgb-flow
:return: None
"""
print(colored('Preparing', 'blue'))
model = Inception3D(weights_path=model_path,
input_shape=(configs.LENGTH, configs.IMG_HEIGHT,
configs.IMG_WIDTH, 2))
# read the steering labels and image path
files = os.listdir(configs.TEST_DIR)
inputs = []
starting_index = 1000
end_index = 1000
if model_type == 'rgb':
for i in range(starting_index, starting_index + configs.LENGTH):
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) +
".jpg")
inputs.append(img)
print(colored('Started', 'blue'))
# Run through all images
for i in range(starting_index + configs.LENGTH + 1,
len(files) - 1 - end_index):
for event in pygame.event.get():
if event.type == pygame.QUIT:
break
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) +
".jpg",
resize=False)
in_frame = cv2.resize(img, (configs.IMG_WIDTH, configs.IMG_HEIGHT))
inputs.pop(0)
inputs.append(in_frame)
prediction = model.model.predict(np.array([inputs]))[0][0]
pygame_loop(label=0, prediction=prediction, img=img)
elif model_type == 'flow':
previous = helper.load_image(configs.TEST_DIR + "frame" +
str(starting_index) + ".jpg")
for i in range(starting_index, starting_index + configs.LENGTH):
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) +
".jpg")
in_frame = cv2.resize(img, (configs.IMG_WIDTH, configs.IMG_HEIGHT))
flow = helper.optical_flow(previous=previous, current=in_frame)
inputs.append(flow)
previous = helper.load_image(configs.TEST_DIR + "frame" +
str(starting_index + configs.LENGTH) +
".jpg")
for i in range(starting_index + configs.LENGTH + 1,
len(files) - 1 - end_index):
for event in pygame.event.get():
if event.type == pygame.QUIT:
break
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) +
".jpg",
resize=False)
in_frame = cv2.resize(img, (configs.IMG_WIDTH, configs.IMG_HEIGHT))
flow = helper.optical_flow(previous, in_frame)
inputs.pop(0)
inputs.append(flow)
input_array = np.array([np.asarray(inputs)])
prediction = model.model.predict(input_array)[0][0]
pygame_loop(label=0, prediction=prediction, img=img)
else:
raise Exception('Sorry, the model type is not recognized')
def test_loop(results_paths):
"""
for visualizing the model with the comma AI
test dataset. The ds doesn't contain training labels.
:param model_path: the path of the trained Keras model
:param model_type: the type of model, rgb, flow or rgb-flow
:return: None
"""
print(colored('Preparing', 'blue'))
# read the steering labels and image path
files = os.listdir(configs.TEST_DIR)
inputs = []
starting_index = 8000
end_index = 1000
for i in range(starting_index, starting_index + configs.LENGTH):
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) + ".jpg")
inputs.append(img)
print(colored('Started', 'blue'))
# Run through all images
for i in range(starting_index + configs.LENGTH + 1,
len(files) - 1 - end_index):
for event in pygame.event.get():
if event.type == pygame.QUIT:
break
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) + ".jpg",
resize=False)
in_frame = cv2.resize(img, (configs.IMG_WIDTH, configs.IMG_HEIGHT))
inputs.pop(0)
inputs.append(in_frame)
pygame_loop(prediction=prediction, img=img)
# ==========================================
previous = helper.load_image(configs.TEST_DIR + "frame" +
str(starting_index) + ".jpg")
for i in range(starting_index, starting_index + configs.LENGTH):
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) + ".jpg")
in_frame = cv2.resize(img, (configs.IMG_WIDTH, configs.IMG_HEIGHT))
flow = helper.optical_flow(previous=previous, current=in_frame)
inputs.append(flow)
previous = helper.load_image(configs.TEST_DIR + "frame" +
str(starting_index + configs.LENGTH) + ".jpg")
for i in range(starting_index + configs.LENGTH + 1,
len(files) - 1 - end_index):
for event in pygame.event.get():
if event.type == pygame.QUIT:
break
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) + ".jpg",
resize=False)
in_frame = cv2.resize(img, (configs.IMG_WIDTH, configs.IMG_HEIGHT))
flow = helper.optical_flow(previous, in_frame)
inputs.pop(0)
inputs.append(flow)
input_array = np.array([np.asarray(inputs)])
prediction = model.model.predict(input_array)[0][0]
pygame_loop(prediction=prediction, img=img)
def visualize_accel(model_path, label_path, model_type):
print(colored('Preparing', 'blue'))
model = Inception3D(weights_path=model_path,
input_shape=(configs.LENGTH, configs.IMG_HEIGHT,
configs.IMG_WIDTH, 3))
# read the steering labels and image path
labels = pd.read_csv(label_path).values
inputs = []
starting_index = 2000
end_index = 0
# init_speed = labels[starting_index][2]
if model_type == 'rgb':
for i in range(starting_index, starting_index + configs.LENGTH):
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) +
".jpg")
inputs.append(img)
print(colored('Started', 'blue'))
# Run through all images
for i in range(starting_index + configs.LENGTH + 1,
len(labels) - 1 - end_index):
for event in pygame.event.get():
if event.type == pygame.QUIT:
break
img = helper.load_image(
"/home/neil/dataset/speedchallenge/data/train/" +
str(labels[i][1]),
resize=False)
in_frame = cv2.resize(img, (configs.IMG_WIDTH, configs.IMG_HEIGHT))
inputs.pop(0)
inputs.append(in_frame)
pred_accel = model.model.predict(np.array([np.asarray(inputs)
]))[0][0]
label_accel = (labels[i][2] - labels[i - 1][2]) * 20
pygame_loop(label=label_accel, prediction=pred_accel, img=img)
elif model_type == 'flow':
previous = helper.load_image(
"/home/neil/dataset/speedchallenge/data/train/" +
str(labels[i][1]))
for i in range(starting_index, starting_index + configs.LENGTH):
img = helper.load_image(configs.TEST_DIR + "frame" + str(i) +
".jpg")
in_frame = cv2.resize(img, (configs.IMG_WIDTH, configs.IMG_HEIGHT))
flow = helper.optical_flow(previous=previous, current=in_frame)
inputs.append(flow)
previous = helper.load_image(
"/home/neil/dataset/speedchallenge/data/train/" +
str(labels[i][1]))
for i in range(starting_index + configs.LENGTH + 1,
len(labels) - 1 - end_index):
for event in pygame.event.get():
if event.type == pygame.QUIT:
break
img = helper.load_image(
"/home/neil/dataset/speedchallenge/data/train/" +
str(labels[i][1]))
in_frame = cv2.resize(img, (configs.IMG_WIDTH, configs.IMG_HEIGHT))
flow = helper.optical_flow(previous, in_frame)
inputs.pop(0)
inputs.append(flow)
pred_accel = model.model.predict(np.array([np.asarray(inputs)
]))[0][0]
label_accel = (labels[i][2] - labels[i - 1][2]) * 20
pygame_loop(label=label_accel, prediction=pred_accel, img=img)
else:
raise Exception('Sorry, the model type is not recognized')