def main():
actions = ["walking", "eating", "smoking", "discussion"]
# TODO make this a runtime option
# Uncomment th eliens
actions.extend([
"directions", "greeting", "phoning", "posing", "purchases", "sitting",
"sittingdown", "takingphoto", "waiting", "walkingdog",
"walkingtogether"
])
# Parameters for dummy model. We only build the model to load the data.
one_hot = False
FLAGS = Object()
FLAGS.data_dir = "../data/h3.6m/dataset"
FLAGS.architecture = "basic"
FLAGS.seq_length_in = 50
FLAGS.seq_length_out = 100
FLAGS.num_layers = 1
FLAGS.size = 128
FLAGS.max_gradient_norm = 5
FLAGS.batch_size = 8
FLAGS.learning_rate = 0.005
FLAGS.learning_rate_decay_factor = 1
summaries_dir = "./log/"
FLAGS.loss_to_use = "sampling_based"
FLAGS.omit_one_hot = True,
FLAGS.residual_velocities = False,
dtype = tf.float32
# Baselines are very simple. No need to use the GPU.
with tf.Session(config=tf.ConfigProto(device_count={"GPU": 0})) as sess:
model = seq2seq_model.Seq2SeqModel(
FLAGS.architecture,
FLAGS.seq_length_in,
FLAGS.seq_length_out,
FLAGS.size, # hidden layer size
FLAGS.num_layers,
FLAGS.max_gradient_norm,
FLAGS.batch_size,
FLAGS.learning_rate,
FLAGS.learning_rate_decay_factor,
summaries_dir,
FLAGS.loss_to_use,
len(actions),
not FLAGS.omit_one_hot,
FLAGS.residual_velocities,
dtype=dtype)
# Load the data
_, test_set, data_mean, data_std, dim_to_ignore, dim_to_use = translate.read_all_data(
actions, FLAGS.seq_length_in, FLAGS.seq_length_out, FLAGS.data_dir,
not FLAGS.omit_one_hot)
# Get all the data, denormalize and convert it to euler angles
poses_data = {}
for action in actions:
enc_in, dec_in, dec_out = model.get_batch_srnn(test_set, action)
enc_in = denormalize_and_convert_to_euler(enc_in, data_mean,
data_std, dim_to_ignore,
actions,
not FLAGS.omit_one_hot)
dec_in = denormalize_and_convert_to_euler(dec_in, data_mean,
data_std, dim_to_ignore,
actions,
not FLAGS.omit_one_hot)
dec_out = denormalize_and_convert_to_euler(dec_out, data_mean,
data_std, dim_to_ignore,
actions,
not FLAGS.omit_one_hot)
poses_data[action] = (enc_in, dec_in, dec_out)
# Compute baseline errors
errs_constant_frame = running_average(poses_data, actions, 1)
# running_average_2 = running_average( poses_data, actions, 2 )
# running_average_4 = running_average( poses_data, actions, 4 )
mean_error = {}
for i in [1, 3, 7, 9, 24]:
mean_error[i] = 0
print()
print("=== Zero-velocity (running avg. 1) ===")
print("{0: <16} | {1:4d} | {2:4d} | {3:4d} | {4:4d} | {5:4d}".format(
"milliseconds", 80, 160, 320, 400, 1000))
for action in actions:
print("{0: <16} | {1:.2f} | {2:.2f} | {3:.2f} | {4:.2f} | {5:.2f}".
format(action, errs_constant_frame[action][1],
errs_constant_frame[action][3],
errs_constant_frame[action][7],
errs_constant_frame[action][9],
errs_constant_frame[action][24]))
for i in [1, 3, 7, 9, 24]:
mean_error[i] += errs_constant_frame[action][i]
print("{0: <16} | {1:.2f} | {2:.2f} | {3:.2f} | {4:.2f} | {5:.2f}".
format("average", mean_error[1] / 15, mean_error[3] / 15,
mean_error[7] / 15, mean_error[9] / 15,
mean_error[24] / 15))
def main():
actions = define_actions( tf.app.flags.FLAGS.action )
# Parameters for dummy model. We only build the model to load the data.
one_hot = False
FLAGS = Object()
FLAGS.data_dir = "./data/h3.6m/dataset"
FLAGS.architecture = "tied"
FLAGS.seq_length_in = 50
FLAGS.seq_length_out = 100
FLAGS.num_layers = 1
FLAGS.size = 128
FLAGS.max_gradient_norm = 5
FLAGS.batch_size = 8
FLAGS.learning_rate = 0.005
FLAGS.learning_rate_decay_factor = 1
summaries_dir = "./log/"
FLAGS.loss_to_use = "sampling_based"
FLAGS.omit_one_hot = True,
FLAGS.residual_velocities = False,
dtype = tf.float32
# Baselines are very simple. No need to use the GPU.
with tf.Session(config=tf.ConfigProto( device_count = {"GPU": 0})) as sess:
model = seq2seq_model.Seq2SeqModel(
FLAGS.architecture,
FLAGS.seq_length_in,
FLAGS.seq_length_out,
FLAGS.size, # hidden layer size
FLAGS.num_layers,
FLAGS.max_gradient_norm,
FLAGS.batch_size,
FLAGS.learning_rate,
FLAGS.learning_rate_decay_factor,
summaries_dir,
FLAGS.loss_to_use,
len( actions ),
not FLAGS.omit_one_hot,
FLAGS.residual_velocities,
dtype=dtype)
# Load the data
_, test_set, data_mean, data_std, dim_to_ignore, dim_to_use = translate.read_all_data(
actions, FLAGS.seq_length_in, FLAGS.seq_length_out, FLAGS.data_dir, not FLAGS.omit_one_hot )
# Get all the data, denormalize and convert it to euler angles
poses_data = {}
for action in actions:
enc_in, dec_in, dec_out = model.get_batch_srnn( test_set, action )
enc_in = denormalize_and_convert_to_euler(
enc_in, data_mean, data_std, dim_to_ignore, actions, not FLAGS.omit_one_hot )
dec_in = denormalize_and_convert_to_euler(
dec_in, data_mean, data_std, dim_to_ignore, actions, not FLAGS.omit_one_hot )
dec_out = denormalize_and_convert_to_euler(
dec_out, data_mean, data_std, dim_to_ignore, actions, not FLAGS.omit_one_hot )
poses_data[action] = (enc_in, dec_in, dec_out)
# Compute baseline errors
errs_constant_frame = running_average( poses_data, actions, 1 )
running_average_2 = running_average( poses_data, actions, 2 )
running_average_4 = running_average( poses_data, actions, 4 )
print()
print("=== Zero-velocity (running avg. 1) ===")
print("{0: <16} | {1:4d} | {2:4d} | {3:4d} | {4:4d}".format("milliseconds", 80, 160, 380, 400))
for action in actions:
print("{0: <16} | {1:.2f} | {2:.2f} | {3:.2f} | {4:.2f}".format( action,
errs_constant_frame[action][1], errs_constant_frame[action][3],
errs_constant_frame[action][7], errs_constant_frame[action][9] ))
print()
print("=== Runnning avg. 2 ===")
print("{0: <16} | {1:4d} | {2:4d} | {3:4d} | {4:4d}".format("milliseconds", 80, 160, 380, 400))
for action in actions:
print("{0: <16} | {1:.2f} | {2:.2f} | {3:.2f} | {4:.2f}".format( action,
running_average_2[action][1], running_average_2[action][3],
running_average_2[action][7], running_average_2[action][9] ))
print()
print("=== Runnning avg. 4 ===")
print("{0: <16} | {1:4d} | {2:4d} | {3:4d} | {4:4d}".format("milliseconds", 80, 160, 380, 400))
for action in actions:
print("{0: <16} | {1:.2f} | {2:.2f} | {3:.2f} | {4:.2f}".format( action,
running_average_4[action][1], running_average_4[action][3],
running_average_4[action][7], running_average_4[action][9] ))