def convert_bag_file(bag_file_path, output_dir, examples_per_file):
print "Processing bag file:", bag_file_path
test_proportion = 0.3
examples = ExampleSet()
bridge = CvBridge()
bag = rosbag.Bag(bag_file_path)
iter_obj = iter(bag.read_messages(topics=["/steering"]))
topic2, msg, t2 = iter_obj.next()
topics = bag.get_type_and_topic_info()[1].keys()
if '/camera/image_rect' in topics:
camera_topic = '/camera/image_rect'
elif '/camera/image_raw' in topics:
camera_topic = '/camera/image_raw'
else:
print "ERROR: camera topic not recognized"
sys.exit(2)
for topic1, img, t1 in bag.read_messages(topics=[camera_topic]):
try:
cv_image = bridge.imgmsg_to_cv2(img, 'bgr8')
# use lower half of picture
halfHeight = cv_image.shape[0] // 2
cv_image = cv_image[halfHeight:, :]
# resize
cv_image = cv2.resize(cv_image, (input_shape[1], input_shape[0]))
while (t1 >= t2):
try:
topic2, msg, t2 = iter_obj.next()
except StopIteration:
print "caught StopIteration for steering message iterator"
break
angle = float(str(msg).split()[-1])
ex = Example(cv_image, angle)
if len(examples.train) >= examples_per_file:
save_next_filename(examples, output_dir)
examples = ExampleSet()
L = examples.test if random.random() < test_proportion else examples.train
L.append(ex)
except CvBridgeError as e:
print e
bag.close()
save_next_filename(examples, output_dir)
def batch_generator(isValidation=False):
gen_epochs = 1 if isValidation else epochs
for epoch in range(gen_epochs):
exampleSetFiles = list(exampleSetFiles_const)
random.shuffle(exampleSetFiles)
while len(exampleSetFiles) > 0:
D = []
while len(exampleSetFiles) > 0 and len(D) < n_examples_to_load:
data = ExampleSet.load(
os.path.join(exampleSetDir, exampleSetFiles.pop()))
D += data.test if isValidation else data.train
if not isValidation: random.shuffle(D)
X = format_inputs(D)
#create output bins
labels = np.array([defineCategory(ex.angle) for ex in D])
if not isValidation:
for i in range(len(X)):
if random.random() < 0.4: # 40% of images are flipped
X[i] = cv2.flip(X[i], 1)
labels[i] = labels[i][::-1]
for i in range(0, len(X), batch_size):
xs = X[i:i + batch_size]
ys = labels[i:i + batch_size]
yield (xs, ys)
def batch_generator(isValidation = False):
gen_epochs = 1 if isValidation else epochs
for epoch in range(gen_epochs):
exampleSetFiles = list(exampleSetFiles_const)
random.shuffle(exampleSetFiles)
while len(exampleSetFiles) > 0:
D = []
while len(exampleSetFiles) > 0 and len(D) < n_examples_to_load:
data = ExampleSet.load(os.path.join(exampleSetDir, exampleSetFiles.pop()))
D += data.test if isValidation else data.train
if not isValidation: random.shuffle(D)
X = format_inputs(D)
# create output bins
labels = np.array([defineCategory(ex.angle) for ex in D])
if not isValidation:
for i in range(len(X)):
if random.random() < 0.4: # 40% of images are flipped
X[i] = cv2.flip(X[i], 1)
labels[i] = labels[i][::-1]
for i in range(0, len(X), batch_size):
xs = X[i: i + batch_size]
ys = labels[i: i + batch_size]
yield (xs, ys)
config = train_profiles[sys.argv[3]]
epochs = config.epochs
categories = config.categories
model = make_model()
model.summary()
exampleSetDir = sys.argv[1]
exampleSetFiles_const = tuple(f for f in os.listdir(exampleSetDir)
if '.pkl.lz4' in f)
n_training_examples = 0
n_test_examples = 0
cnt = collections.Counter()
for f in exampleSetFiles_const:
data = ExampleSet.load(os.path.join(exampleSetDir, f))
n_training_examples += len(data.train)
n_test_examples += len(data.test)
for ex in data.train:
i = np.argmax(defineCategory(ex.angle))
cnt[i] += 1
print "total training examples:", n_training_examples
print "training label counts:", cnt
def batch_generator(isValidation=False):
gen_epochs = 1 if isValidation else epochs
for epoch in range(gen_epochs):
exampleSetFiles = list(exampleSetFiles_const)
random.shuffle(exampleSetFiles)
def main():
global categories, n_examples_to_load, batch_size
rospy.init_node("nn_training")
startTime = time.time()
model_path = rospy.get_param("~model_output_path")
exampleSetDir = rospy.get_param("~example_set_dir")
epochs = int(rospy.get_param("~epochs"))
categories = rospy.get_param("~positive_nonzero_categories")
categories = string.strip(categories).split(" ")
categories = [float(x) for x in categories]
categories = expand_categories(categories)
model = make_model()
model.summary()
exampleSetFiles_const = tuple(f for f in os.listdir(exampleSetDir)
if '.pkl.lz4' in f)
n_training_examples = 0
n_test_examples = 0
cnt = collections.Counter()
for f in exampleSetFiles_const:
data = ExampleSet.load(os.path.join(exampleSetDir, f))
n_training_examples += len(data.train)
n_test_examples += len(data.test)
for ex in data.train:
i = np.argmax(defineCategory(ex.angle))
cnt[i] += 1
print "total training examples:", n_training_examples
print "training label counts:", cnt
def batch_generator(isValidation=False):
gen_epochs = 1 if isValidation else epochs
for epoch in range(gen_epochs):
exampleSetFiles = list(exampleSetFiles_const)
random.shuffle(exampleSetFiles)
while len(exampleSetFiles) > 0:
D = []
while len(exampleSetFiles) > 0 and len(D) < n_examples_to_load:
data = ExampleSet.load(
os.path.join(exampleSetDir, exampleSetFiles.pop()))
D += data.test if isValidation else data.train
if not isValidation: random.shuffle(D)
X = format_inputs(D)
# create output bins
labels = np.array([defineCategory(ex.angle) for ex in D])
if not isValidation:
for i in range(len(X)):
if random.random() < 0.4: # 40% of images are flipped
X[i] = cv2.flip(X[i], 1)
labels[i] = labels[i][::-1]
for i in range(0, len(X), batch_size):
xs = X[i:i + batch_size]
ys = labels[i:i + batch_size]
yield (xs, ys)
try:
n_minibatches = int(math.ceil(float(n_training_examples) / batch_size))
model.fit_generator(batch_generator(),
steps_per_epoch=n_minibatches,
epochs=epochs,
verbose=1)
print "elapsed time:", time.time() - startTime
n_minibatches = int(math.ceil(float(n_test_examples) / batch_size))
loss, acc = model.evaluate_generator(batch_generator(True),
steps=n_minibatches)
print "validation loss:", loss, "| validation accuracy:", acc
finally:
model.save(model_path)
print "\nsaved model to", model_path
def main():
global categories, n_examples_to_load, batch_size
rospy.init_node("nn_training")
startTime = time.time()
model_path = rospy.get_param("~model_output_path", None)
exampleSetDir = rospy.get_param("~example_set_dir", None)
epochs = int(rospy.get_param("~epochs", None))
categories = rospy.get_param("~positive_nonzero_categories", None)
categories = string.strip(categories).split(" ")
categories = [float(x) for x in categories]
categories = expand_categories(categories)
model = make_model()
model.summary()
exampleSetFiles_const = tuple(f for f in os.listdir(exampleSetDir) if '.pkl.lz4' in f)
n_training_examples = 0
n_test_examples = 0
cnt = collections.Counter()
for f in exampleSetFiles_const:
data = ExampleSet.load(os.path.join(exampleSetDir, f))
n_training_examples += len(data.train)
n_test_examples += len(data.test)
for ex in data.train:
i = np.argmax(defineCategory(ex.angle))
cnt[i] += 1
print "total training examples:", n_training_examples
print "training label counts:", cnt
def batch_generator(isValidation = False):
gen_epochs = 1 if isValidation else epochs
for epoch in range(gen_epochs):
exampleSetFiles = list(exampleSetFiles_const)
random.shuffle(exampleSetFiles)
while len(exampleSetFiles) > 0:
D = []
while len(exampleSetFiles) > 0 and len(D) < n_examples_to_load:
data = ExampleSet.load(os.path.join(exampleSetDir, exampleSetFiles.pop()))
D += data.test if isValidation else data.train
if not isValidation: random.shuffle(D)
X = format_inputs(D)
# create output bins
labels = np.array([defineCategory(ex.angle) for ex in D])
if not isValidation:
for i in range(len(X)):
if random.random() < 0.4: # 40% of images are flipped
X[i] = cv2.flip(X[i], 1)
labels[i] = labels[i][::-1]
for i in range(0, len(X), batch_size):
xs = X[i: i + batch_size]
ys = labels[i: i + batch_size]
yield (xs, ys)
try:
n_minibatches = int(math.ceil(float(n_training_examples) / batch_size))
model.fit_generator(batch_generator(),
steps_per_epoch=n_minibatches,
epochs=epochs,
verbose=1)
print "elapsed time:", time.time() - startTime
n_minibatches = int(math.ceil(float(n_test_examples) / batch_size))
loss, acc = model.evaluate_generator(batch_generator(True), steps=n_minibatches)
print "validation loss:", loss, "| validation accuracy:", acc
finally:
model.save(model_path)
print "\nsaved model to", model_path