def loss_car_stop(logits, net_outputs, batch_size=None):
# net_outputs contains is_stop
labels = net_outputs[0] # shape: N * F
# reshape to 1 dimension
labels = tf.reshape(labels, [-1])
prediction = logits[0] # shape: (N * F) * 2
# filter the no ground truth data
labels, prediction = util.filter_no_groundtruth_label(labels, prediction)
labels_shape = tf.shape(labels)
effective_batch_size = labels_shape[0]
num_classes = prediction.get_shape()[-1].value
dense_labels = util.to_one_hot_label(labels, effective_batch_size,
num_classes)
if FLAGS.class_balance_path != "":
path = FLAGS.class_balance_path + "_stop.npy"
empirical_distribution = np.load(path)
weights = util.loss_weights(empirical_distribution,
FLAGS.class_balance_epsilon)
print("using weighted training: ", weights)
mask = tf.gather(weights, labels)
else:
mask = 1.0
# Cross entropy loss for the main softmax prediction.
slim.losses.softmax_cross_entropy(prediction, dense_labels, weight=mask)
def loss_car_joint(logits, net_outputs, batch_size=None):
# net_outputs contains is_stop, turn, locs
future_labels = net_outputs[2] # shape: N * F * 2
# reshape to 2 dimension
num_classes = future_labels.get_shape()[-1].value
future_labels = tf.reshape(future_labels, [-1, num_classes])
dense_labels = tf.py_func(course_speed_to_joint_bin, [future_labels],
[tf.float32])
if FLAGS.class_balance_path != "":
path = FLAGS.class_balance_path + "_joint.npy"
dist = np.load(path)
weights = util.loss_weights(dist, FLAGS.class_balance_epsilon)
print("using weighted training: ", weights)
# assume the label being the max response at that point
labels = tf.argmax(dense_labels, dimension=1)
masks = tf.gather(weights, labels)
else:
masks = 1.0
future_predict = logits[0]
slim.losses.softmax_cross_entropy(future_predict,
dense_labels,
weight=masks)
def loss_car_discrete(logits, net_outputs, batch_size=None):
# net_outputs contains is_stop, turn
dense_labels = net_outputs[1] # shape: N * F * nclass
# reshape to 2 dimension
num_classes = dense_labels.get_shape()[-1].value
dense_labels = tf.reshape(dense_labels, [-1, num_classes])
if FLAGS.class_balance_path != "":
path = FLAGS.class_balance_path + "_discrete.npy"
empirical_distribution = np.load(path)
weights = util.loss_weights(empirical_distribution,
FLAGS.class_balance_epsilon)
print("using weighted training: ", weights)
# assume the label being the max response at that point
labels = tf.argmax(dense_labels, dimension=1)
mask = tf.gather(weights, labels)
else:
mask = 1.0
# Cross entropy loss for the main softmax prediction.
slim.losses.softmax_cross_entropy(logits[0], dense_labels, weight=mask)
def loss_car_loc_xy(logits, net_outputs, batch_size=None):
# net_outputs contains is_stop, turn, locs
future_labels = net_outputs[2] # shape: N * F * 2
# reshape to 2 dimension
num_classes = future_labels.get_shape()[-1].value
NF = future_labels.get_shape()[0].value * \
future_labels.get_shape()[1].value
future_labels = tf.reshape(future_labels, [-1, num_classes])
dense_course, dense_speed = tf.py_func(call_label_to_dense_smooth,
[future_labels],
[tf.float32, tf.float32])
if FLAGS.class_balance_path != "":
path = FLAGS.class_balance_path + "_continuous.npy"
dists = np.load(path)
masks = []
dense_labels = [dense_course, dense_speed]
for i in range(2):
weights = util.loss_weights(dists[i], FLAGS.class_balance_epsilon)
print("using weighted training: ", weights)
# assume the label being the max response at that point
labels = tf.argmax(dense_labels[i], dimension=1)
mask = tf.gather(weights, labels)
mask.set_shape([NF])
masks.append(mask)
else:
masks = [1.0, 1.0]
future_predict = logits[0] # shape: (N*F) * 2Nbins
n = FLAGS.discretize_n_bins
slim.losses.softmax_cross_entropy(future_predict[:, 0:n],
dense_course,
scope="cross_entropy_loss/course",
weight=masks[0])
slim.losses.softmax_cross_entropy(future_predict[:, n:],
dense_speed,
scope="cross_entropy_loss/speed",
weight=masks[1])