def __init__(self,
model_config_name,
model_path,
truncate_len=20,
batch_size=1):
# currently, we use a sliding window fashion for evaluation, that's inefficient but convenient to implement
self.truncate_len = truncate_len
self.latest_frames = []
for _ in range(truncate_len):
self.latest_frames.append(np.zeros((IMSZ, IMSZ, 3),
dtype=np.uint8))
# call the config.py for setup
import config
common_config("eval")
config_fun = getattr(config, model_config_name)
config_fun("eval")
common_config_post("eval")
# Tensors in has the format: [images, speed] for basic usage, excluding only_seg
# For now, we decide not to support previous speed as input, thus we use a fake speed (-1) now
# and ensures the speed is not used by asserting FLAGS.use_previous_speed_feature==False
assert (not hasattr(FLAGS, "use_previous_speed_feature")) or (
FLAGS.use_previous_speed_feature == False)
# batch size 1 all the time, length undetermined, width and height are IMSZ
FLAGS.__setattr__('log_device_placement', True)
self.tensors_in = tf.placeholder(tf.uint8,
shape=(batch_size, truncate_len, IMSZ,
IMSZ, 3),
name="images_input")
self.speed = None
logits_all = model.inference([self.tensors_in, self.speed],
-1,
for_training=False)
# Restore the moving average version of the learned variables for eval.
variable_averages = tf.train.ExponentialMovingAverage(
model.MOVING_AVERAGE_DECAY)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
config = tf.ConfigProto(intra_op_parallelism_threads=1)
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
saver.restore(self.sess, model_path)
self.logits = tf.nn.softmax(logits_all[0])
init_op = tf.initialize_local_variables()
self.sess.run(init_op)
FLAGS = tf.app.flags.FLAGS
flags_passthrough = FLAGS._parse_flags()
from config import common_config, common_config_post
IMSZ = 228
model_config_name = "discrete_fcn_lstm"
truncate_len = 20
#sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
#write_dict_to_file('./results/flags_orig.txt', FLAGS.__flags)
import config
common_config("eval")
config_fun = getattr(config, model_config_name)
config_fun("eval")
common_config_post("eval")
#write_dict_to_file('./results/flags_altered.txt', FLAGS.__flags)
# Tensors in has the format: [images, speed] for basic usage, excluding only_seg
# For now, we decide not to support previous speed as input, thus we use a fake speed (-1) now
# and ensures the speed is not used by asserting FLAGS.use_previous_speed_feature==False
assert (not hasattr(FLAGS, "use_previous_speed_feature")) or (
FLAGS.use_previous_speed_feature == False)
# batch size 1 all the time, length undetermined, width and height are IMSZ
#FLAGS.__setattr__('log_device_placement', True)
tensors_in = tf.placeholder(tf.uint8,
shape=(256, truncate_len, IMSZ, IMSZ, 3),
def __init__(self, model_config_name, model_path, truncate_len=20, config_name="config", config_path=".", is_lstm=False):
self.is_lstm = is_lstm
if is_lstm:
assert truncate_len==1, \
"using lstm should set truncate_len to 1, otherwise waste of computing resource"
# currently, we use a sliding window fashion for evaluation, that's inefficient but convenient to implement
self.truncate_len = truncate_len
self.latest_frames = []
for _ in range(truncate_len):
self.latest_frames.append(np.zeros((IMSZ, IMSZ, 3), dtype=np.uint8))
# call the config.py for setup
sys.path.append(config_path)
config = importlib.import_module(config_name)
common_config("eval")
config_fun = getattr(config, model_config_name)
config_fun("eval")
common_config_post("eval")
# Tensors in has the format: [images, speed] for basic usage, excluding only_seg
# For now, we decide not to support previous speed as input, thus we use a fake speed (-1) now
# and ensures the speed is not used by asserting FLAGS.use_previous_speed_feature==False
assert (not hasattr(FLAGS, "use_previous_speed_feature")) or (FLAGS.use_previous_speed_feature == False)
# batch size 1 all the time, length undetermined, width and height are IMSZ
self.tensors_in = tf.placeholder(tf.uint8, shape=(1, truncate_len, IMSZ, IMSZ, 3), name="images_input")
self.speed = None
if is_lstm:
self.initial_state = ((tf.placeholder(tf.float32,
shape=(1, int(FLAGS.lstm_hidden_units)),
name="state_placeholder1"),
tf.placeholder(tf.float32,
shape=(1, int(FLAGS.lstm_hidden_units)),
name="state_placeholder2")), )
FLAGS.phase = "rnn_inference"
else:
self.initial_state = None
logits_all = model.inference([self.tensors_in, self.speed], -1, for_training=False,
initial_state=self.initial_state)
# Restore the moving average version of the learned variables for eval.
variable_averages = tf.train.ExponentialMovingAverage(model.MOVING_AVERAGE_DECAY)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
config = tf.ConfigProto(intra_op_parallelism_threads=1)
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
saver.restore(self.sess, model_path)
self.logits = logits_all[0]
if is_lstm:
self.state_tensor = logits_all[-1]
self.state_value=[[np.zeros((1, int(FLAGS.lstm_hidden_units)), dtype=np.float32),
np.zeros((1, int(FLAGS.lstm_hidden_units)), dtype=np.float32)]]
init_op = tf.initialize_local_variables()
self.sess.run(init_op)