def test():
""" Evaluate on test set """
actions = data_utils.define_actions(FLAGS.action)
number_of_actions = len(actions)
# Load camera parameters
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14)
# Read stacked hourglass 2D predictions if use_sh, otherwise use groundtruth 2D projections
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(
actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data(
actions, FLAGS.data_dir, rcams)
print("done reading and normalizing data.")
# Avoid using the GPU if requested
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(device_count=device_count,
allow_soft_placement=True)) as sess:
# === Create the model ===
print("Creating %d bi-layers of %d units." %
(FLAGS.num_layers, FLAGS.linear_size))
model = create_model(sess, actions, FLAGS.batch_size)
model.train_writer.add_graph(sess.graph)
print("Model created")
#=== This is the training loop ===
step_time, loss, val_loss = 0.0, 0.0, 0.0
current_step = 0 if FLAGS.load <= 0 else FLAGS.load + 1
previous_losses = []
step_time, loss = 0, 0
current_epoch = 0
log_every_n_batches = 100
for _ in xrange(1):
# === Testing after this epoch ===
isTraining = False
if FLAGS.evaluateActionWise:
print("{0:=^12} {1:=^6}".format(
"Action", "mm")) # line of 30 equal signs
cum_err = 0
for action in actions:
print("{0:<12} ".format(action), end="")
# Get 2d and 3d testing data for this action
action_test_set_2d = get_action_subset(test_set_2d, action)
action_test_set_3d = get_action_subset(test_set_3d, action)
encoder_inputs, decoder_outputs, _ = data_utils.get_all_batches( action_test_set_2d, action_test_set_3d, FLAGS.camera_frame, training=False,\
n_context=FLAGS.n_context, new_dim=False, batch_size=FLAGS.batch_size)
act_err, _, step_time, loss = evaluate_batches(
sess,
model,
data_mean_3d,
data_std_3d,
dim_to_use_3d,
dim_to_ignore_3d,
data_mean_2d,
data_std_2d,
dim_to_use_2d,
dim_to_ignore_2d,
current_step,
encoder_inputs,
decoder_outputs,
test_set_2d=action_test_set_2d)
cum_err = cum_err + act_err
print("{0:>6.2f}".format(act_err))
# summaries = sess.run( model.err_mm_summary, {model.err_mm: float(cum_err/float(len(actions)))} )
# model.test_writer.add_summary( summaries, current_step )
print("{0:<12} {1:>6.2f}".format("Average", cum_err /
float(len(actions))))
print("{0:=^19}".format(''))
else:
pass
# n_joints = 17 if not(FLAGS.predict_14) else 14
# encoder_inputs, decoder_outputs, _ = data_utils.get_all_batches( test_set_2d, test_set_3d, FLAGS.camera_frame,\
# training=False, n_context=FLAGS.n_context, new_dim=False, batch_size=FLAGS.batch_size)
#
# total_err, joint_err, step_time, loss = evaluate_batches( sess, model,
# data_mean_3d, data_std_3d, dim_to_use_3d, dim_to_ignore_3d,
# data_mean_2d, data_std_2d, dim_to_use_2d, dim_to_ignore_2d,
# current_step, encoder_inputs, decoder_outputs, current_epoch )
#
# print("=============================\n"
# "Step-time (ms): %.4f\n"
# "Val loss avg: %.4f\n"
# "Val error avg (mm): %.2f\n"
# "=============================" % ( 1000*step_time, loss, total_err ))
#
# for i in range(n_joints):
# # 6 spaces, right-aligned, 5 decimal places
# print("Error in joint {0:02d} (mm): {1:>5.2f}".format(i+1, joint_err[i]))
# print("=============================")
# Log the error to tensorboard
# summaries = sess.run( model.err_mm_summary, {model.err_mm: total_err} )
# model.test_writer.add_summary( summaries, current_step )
# Save the model
# print( "Saving the model... ", end="" )
# start_time = time.time()
# model.saver.save(sess, os.path.join(train_dir, 'checkpoint'), global_step=current_step )
# print( "done in {0:.2f} ms".format(1000*(time.time() - start_time)) )
# Reset global time and loss
step_time, loss = 0, 0
sys.stdout.flush()
def train():
"""Train a linear model for 3d pose estimation"""
actions = data_utils.define_actions( FLAGS.action )
number_of_actions = len( actions )
# Load camera parameters
SUBJECT_IDS = [1,5,6,7,8,9,11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14 )
# Read stacked hourglass 2D predictions if use_sh, otherwise use groundtruth 2D projections
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data( actions, FLAGS.data_dir, rcams )
print( "done reading and normalizing data." )
# Avoid using the GPU if requested
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(
device_count=device_count,
allow_soft_placement=True )) as sess:
# === Create the model ===
print("Creating %d bi-layers of %d units." % (FLAGS.num_layers, FLAGS.linear_size))
model = create_model( sess, actions, FLAGS.batch_size )
model.train_writer.add_graph( sess.graph )
print("Model created")
#=== This is the training loop ===
step_time, loss, val_loss = 0.0, 0.0, 0.0
current_step = 0 if FLAGS.load <= 0 else FLAGS.load + 1
previous_losses = []
step_time, loss = 0, 0
current_epoch = 0
log_every_n_batches = 100
for _ in xrange( FLAGS.epochs ):
current_epoch = current_epoch + 1
# === Load training batches for one epoch ===
encoder_inputs, decoder_outputs = model.get_all_batches( train_set_2d, train_set_3d, FLAGS.camera_frame, training=True )
nbatches = len( encoder_inputs )
print("There are {0} train batches".format( nbatches ))
start_time, loss = time.time(), 0.
# === Loop through all the training batches ===
for i in range( nbatches ):
if (i+1) % log_every_n_batches == 0:
# Print progress every log_every_n_batches batches
print("Working on epoch {0}, batch {1} / {2}... ".format( current_epoch, i+1, nbatches), end="" )
enc_in, dec_out = encoder_inputs[i], decoder_outputs[i]
step_loss, loss_summary, lr_summary, _ = model.step( sess, enc_in, dec_out, FLAGS.dropout, isTraining=True )
if (i+1) % log_every_n_batches == 0:
# Log and print progress every log_every_n_batches batches
model.train_writer.add_summary( loss_summary, current_step )
model.train_writer.add_summary( lr_summary, current_step )
step_time = (time.time() - start_time)
start_time = time.time()
print("done in {0:.2f} ms".format( 1000*step_time / log_every_n_batches ) )
loss += step_loss
current_step += 1
# === end looping through training batches ===
loss = loss / nbatches
print("=============================\n"
"Global step: %d\n"
"Learning rate: %.2e\n"
"Train loss avg: %.4f\n"
"=============================" % (model.global_step.eval(),
model.learning_rate.eval(), loss) )
# === End training for an epoch ===
# === Testing after this epoch ===
isTraining = False
if FLAGS.evaluateActionWise:
print("{0:=^12} {1:=^6}".format("Action", "mm")) # line of 30 equal signs
cum_err = 0
for action in actions:
print("{0:<12} ".format(action), end="")
# Get 2d and 3d testing data for this action
action_test_set_2d = get_action_subset( test_set_2d, action )
action_test_set_3d = get_action_subset( test_set_3d, action )
encoder_inputs, decoder_outputs = model.get_all_batches( action_test_set_2d, action_test_set_3d, FLAGS.camera_frame, training=False)
act_err, _, step_time, loss = evaluate_batches( sess, model,
data_mean_3d, data_std_3d, dim_to_use_3d, dim_to_ignore_3d,
data_mean_2d, data_std_2d, dim_to_use_2d, dim_to_ignore_2d,
current_step, encoder_inputs, decoder_outputs )
cum_err = cum_err + act_err
print("{0:>6.2f}".format(act_err))
summaries = sess.run( model.err_mm_summary, {model.err_mm: float(cum_err/float(len(actions)))} )
model.test_writer.add_summary( summaries, current_step )
print("{0:<12} {1:>6.2f}".format("Average", cum_err/float(len(actions) )))
print("{0:=^19}".format(''))
else:
n_joints = 17 if not(FLAGS.predict_14) else 14
encoder_inputs, decoder_outputs = model.get_all_batches( test_set_2d, test_set_3d, FLAGS.camera_frame, training=False)
total_err, joint_err, step_time, loss = evaluate_batches( sess, model,
data_mean_3d, data_std_3d, dim_to_use_3d, dim_to_ignore_3d,
data_mean_2d, data_std_2d, dim_to_use_2d, dim_to_ignore_2d,
current_step, encoder_inputs, decoder_outputs, current_epoch )
print("=============================\n"
"Step-time (ms): %.4f\n"
"Val loss avg: %.4f\n"
"Val error avg (mm): %.2f\n"
"=============================" % ( 1000*step_time, loss, total_err ))
for i in range(n_joints):
# 6 spaces, right-aligned, 5 decimal places
print("Error in joint {0:02d} (mm): {1:>5.2f}".format(i+1, joint_err[i]))
print("=============================")
# Log the error to tensorboard
summaries = sess.run( model.err_mm_summary, {model.err_mm: total_err} )
model.test_writer.add_summary( summaries, current_step )
# Save the model
print( "Saving the model... ", end="" )
start_time = time.time()
model.saver.save(sess, os.path.join(train_dir, 'checkpoint'), global_step=current_step )
print( "done in {0:.2f} ms".format(1000*(time.time() - start_time)) )
# Reset global time and loss
step_time, loss = 0, 0
sys.stdout.flush()
def hankgogo(gogodata, gogodatafake):
"""Get samples from a model and visualize them"""
actions = data_utils.define_actions(FLAGS.action)
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14)
#if FLAGS.use_sh:
# train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(actions, FLAGS.data_dir)
#else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data(
actions, FLAGS.data_dir, rcams)
print("done reading and normalizing data.")
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(device_count=device_count)) as sess:
# === Create the model ===
print("Creating %d layers of %d units." %
(FLAGS.num_layers, FLAGS.linear_size))
batch_size = 1
model = create_model_my(sess, actions, batch_size)
print("Model loaded")
# Dropout probability 0 (keep probability 1) for sampling
dp = 1.0
poses3d = model.step(sess, gogodata, isTraining=False)
tesmp = poses3d
poses3d = data_utils.unNormalizeData(poses3d, data_mean_3d,
data_std_3d, dim_to_ignore_3d)
model.saver.save(sess, os.path.join(mysave_dir, "gogo"))
# Grab a random batch to visualize
# enc_in, dec_out, poses3d = map( np.vstack, [enc_in, dec_out, poses3d] )
# idx = np.random.permutation( enc_in.shape[0] )
# enc_in, dec_out, poses3d = enc_in[idx, :], dec_out[idx, :], poses3d[idx, :]
# Visualize random samples
import matplotlib.gridspec as gridspec
# 1080p = 1,920 x 1,080
fig = plt.figure(figsize=(19.2, 10.8))
gs1 = gridspec.GridSpec(5, 9) # 5 rows, 9 columns
gs1.update(wspace=-0.00, hspace=0.05) # set the spacing between axes.
plt.axis('off')
subplot_idx, exidx = 1, 1
nsamples = 1
# Plot 2d pose
#ax1 = plt.subplot(gs1[subplot_idx-1])
#p2d = enc_in[exidx,:]
#viz.show2Dpose( p2d, ax1 )
#ax1.invert_yaxis()
# Plot 3d gt
#ax2 = plt.subplot(gs1[subplot_idx], projection='3d')
#p3d = dec_out[exidx,:]
#viz.show3Dpose( p3d, ax2 )
# Plot 3d predictions
ax3 = plt.subplot(gs1[subplot_idx + 1], projection='3d')
p3d = poses3d
viz.show3Dpose(p3d, ax3, lcolor="#9b59b6", rcolor="#2ecc71")
exidx = exidx + 1
subplot_idx = subplot_idx + 3
plt.show()
def main(_):
actions_all = data_utils.define_actions("All")
actions = data_utils.define_actions("Discussion")
# Load camera parameters
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14)
train_set_3d = data_utils.remove_first_frame(train_set_3d)
test_set_3d = data_utils.remove_first_frame(test_set_3d)
train_root_positions = data_utils.remove_first_frame(train_root_positions)
test_root_positions = data_utils.remove_first_frame(test_root_positions)
print("Finished Read 3D Data")
# train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(actions_all, FLAGS.data_dir)
# train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.transform_to_2d_biframe_prediction(train_set_2d,
# test_set_2d,
# data_mean_2d,
# data_std_2d,
# dim_to_ignore_2d,
# dim_to_use_2d)
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data(
actions_all, FLAGS.data_dir, rcams)
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.transform_to_2d_biframe_prediction(
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d,
dim_to_use_2d)
SH_TO_GT_PERM = np.array(
[SH_NAMES.index(h) for h in H36M_NAMES if h != '' and h in SH_NAMES])
assert np.all(SH_TO_GT_PERM == np.array(
[6, 2, 1, 0, 3, 4, 5, 7, 8, 9, 13, 14, 15, 12, 11, 10]))
test_set = {}
manipulation_dir = os.path.dirname(FLAGS.data_dir)
manipulation_dir = os.path.dirname(manipulation_dir)
manipulation_dir += '/manipulation_video/'
manipulation_folders = glob.glob(manipulation_dir + '*')
subj = 1
action = 'manipulation-video'
for folder in manipulation_folders:
seqname = os.path.basename(folder)
with h5py.File(folder + '/' + seqname + '.h5', 'r') as h5f:
poses = h5f['poses'][:]
# Permute the loaded data to make it compatible with H36M
poses = poses[:, SH_TO_GT_PERM, :]
# Reshape into n x (32*2) matrix
poses = np.reshape(poses, [poses.shape[0], -1])
poses_final = np.zeros([poses.shape[0], len(H36M_NAMES) * 2])
dim_to_use_x = np.where(
np.array([x != '' and x != 'Neck/Nose'
for x in H36M_NAMES]))[0] * 2
dim_to_use_y = dim_to_use_x + 1
dim_to_use = np.zeros(len(SH_NAMES) * 2, dtype=np.int32)
dim_to_use[0::2] = dim_to_use_x
dim_to_use[1::2] = dim_to_use_y
poses_final[:, dim_to_use] = poses
print(seqname, poses_final.shape)
poses_final[poses_final == 0.] = 0.1
test_set[(subj, action, seqname)] = poses_final
test_set = data_utils.uni_frame_to_bi_frame(test_set)
test_set_2d = data_utils.normalize_data(test_set, data_mean_2d,
data_std_2d, dim_to_use_2d)
for key in test_set.keys():
test_set[key] = test_set[key][0::2, :]
dim_to_use_12_manipulation_joints = np.array([
3, 4, 5, 6, 7, 8, 9, 10, 11, 18, 19, 20, 21, 22, 23, 24, 25, 26, 51,
52, 53, 54, 55, 56, 57, 58, 59, 75, 76, 77, 78, 79, 80, 81, 82, 83
])
print("Finished Normalize Manipualtion Videos")
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(device_count=device_count)) as sess:
# === Create the model ===
print("Creating %d layers of %d units." %
(FLAGS.num_layers, FLAGS.linear_size))
batch_size = FLAGS.batch_size #Intial code is 64*2
model = predict_3dpose_biframe.create_model(sess, actions_all,
batch_size)
print("Model loaded")
j = 0
for key2d in test_set_2d.keys():
(subj, b, fname) = key2d
# if fname != specific_seqname + '.h5':
# continue
print("Subject: {}, action: {}, fname: {}".format(subj, b, fname))
enc_in = test_set_2d[key2d]
n2d, _ = enc_in.shape
# Split into about-same-size batches
enc_in = np.array_split(enc_in, n2d // 1)
all_poses_3d = []
for bidx in range(len(enc_in)):
# Dropout probability 0 (keep probability 1) for sampling
dp = 1.0
anything = np.zeros((enc_in[bidx].shape[0], 48))
_, _, poses3d = model.step(sess,
enc_in[bidx],
anything,
dp,
isTraining=False)
# Denormalize
enc_in[bidx] = data_utils.unNormalizeData(
enc_in[bidx], data_mean_2d, data_std_2d, dim_to_ignore_2d)
poses3d = data_utils.unNormalizeData(poses3d, data_mean_3d,
data_std_3d,
dim_to_ignore_3d)
all_poses_3d.append(poses3d)
# Put all the poses together
enc_in, poses3d = map(np.vstack, [enc_in, all_poses_3d])
enc_in, poses3d = map(np.vstack, [enc_in, poses3d])
poses3d_12_manipulation = poses3d[:,
dim_to_use_12_manipulation_joints]
annotated_images = glob.glob(manipulation_dir + fname +
'/info/*.xml')
annotated_images = sorted(annotated_images)
# 1080p = 1,920 x 1,080
fig = plt.figure(j, figsize=(10, 10))
gs1 = gridspec.GridSpec(3, 3)
gs1.update(wspace=-0, hspace=0.1) # set the spacing between axes.
plt.axis('off')
subplot_idx = 1
nsamples = 3
for i in np.arange(nsamples):
# Plot 2d Detection
ax1 = plt.subplot(gs1[subplot_idx - 1])
img = mpimg.imread(
manipulation_dir + fname + '/skeleton_cropped/' +
os.path.basename(annotated_images[i]).split('_')[0] +
'.jpg')
ax1.imshow(img)
# Plot 2d pose
ax2 = plt.subplot(gs1[subplot_idx])
# p2d = enc_in[i,:]
# viz.show2Dpose( p2d, ax2 )
# ax2.invert_yaxis()
ax2.imshow(img)
# Plot 3d predictions
# Compute first the procrustion and print error
gt = getJ3dPosFromXML(annotated_images[i])
A = poses3d_12_manipulation[i, :].reshape(gt.shape)
_, Z, T, b, c = procrustes.compute_similarity_transform(
gt, A, compute_optimal_scale=True)
sqerr = np.sqrt(np.sum((gt - (b * A.dot(T)) - c)**2, axis=1))
print("{0} - {1} - Mean Error (mm) : {2}".format(
fname, os.path.basename(annotated_images[i]),
np.mean(sqerr)))
ax3 = plt.subplot(gs1[subplot_idx + 1], projection='3d')
temp = poses3d[i, :].reshape((32, 3))
temp = c + temp.dot(T) #Do not scale
# p3d = temp.reshape((1, 96))
p3d = poses3d[i, :]
viz.show3Dpose(p3d, ax3, lcolor="#9b59b6", rcolor="#2ecc71")
ax3.invert_zaxis()
ax3.invert_yaxis()
subplot_idx = subplot_idx + 3
plt.show()
j += 1
def test():
actions = data_utils.define_actions(FLAGS.action)
# Load camera parameters
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14)
# Read stacked hourglass 2D predictions if use_sh, otherwise use groundtruth 2D projections
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(
actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data(
actions, FLAGS.data_dir, rcams)
# Avoid using the GPU if requested
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(device_count=device_count,
allow_soft_placement=True)) as sess:
# === Create the model ===
print("Creating %d bi-layers of %d units." %
(FLAGS.num_layers, FLAGS.linear_size))
model = create_model(sess, actions, FLAGS.batch_size)
model.train_writer.add_graph(sess.graph)
current_step = 0 if FLAGS.load <= 0 else FLAGS.load + 1
if FLAGS.evaluateActionWise:
logger.info("{0:=^12} {1:=^6}".format(
"Action", "mm")) # line of 30 equal signs
cum_err = 0 # select the mixture model which has mininum error
for action in actions:
# Get 2d and 3d testing data for this action
action_test_set_2d = get_action_subset(test_set_2d, action)
action_test_set_3d = get_action_subset(test_set_3d, action)
encoder_inputs, decoder_outputs, repro_info = model.get_all_batches(
action_test_set_2d,
action_test_set_3d,
FLAGS.camera_frame,
training=False)
act_err, step_time, loss = evaluate_batches(
sess, model, data_mean_3d, data_std_3d, dim_to_use_3d,
dim_to_ignore_3d, data_mean_2d, data_std_2d, dim_to_use_2d,
dim_to_ignore_2d, current_step, encoder_inputs,
decoder_outputs)
cum_err = cum_err + act_err
logger.info('{0:<12} {1:>6.2f}'.format(action, act_err))
summaries = sess.run(
model.err_mm_summary,
{model.err_mm: float(cum_err / float(len(actions)))})
model.test_writer.add_summary(summaries, current_step)
logger.info('{0:<12} {1:>6.2f}'.format(
"Average", cum_err / float(len(actions))))
logger.info('{0:=^19}'.format(''))
def sample():
"""Get samples from a model and visualize them"""
path = '{}/samples_sh'.format(FLAGS.train_dir)
if not os.path.exists(path):
os.makedirs(path)
actions = data_utils.define_actions(FLAGS.action)
# Load camera parameters
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
n_joints = 17 if not (FLAGS.predict_14) else 14
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14)
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(
actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d, _ = data_utils.create_2d_data(
actions, FLAGS.data_dir, rcams)
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(device_count=device_count)) as sess:
# === Create the model ===
batch_size = 128
model = create_model(sess, actions, batch_size)
print("Model loaded")
for key2d in test_set_2d.keys():
(subj, b, fname) = key2d
# choose SittingDown action to visualize
if b == 'SittingDown':
print("Subject: {}, action: {}, fname: {}".format(
subj, b, fname))
# keys should be the same if 3d is in camera coordinates
key3d = key2d if FLAGS.camera_frame else (
subj, b, '{0}.h5'.format(fname.split('.')[0]))
key3d = (subj, b, fname[:-3]) if (
fname.endswith('-sh')) and FLAGS.camera_frame else key3d
enc_in = test_set_2d[key2d]
n2d, _ = enc_in.shape
dec_out = test_set_3d[key3d]
n3d, _ = dec_out.shape
assert n2d == n3d
# Split into about-same-size batches
enc_in = np.array_split(enc_in, n2d // batch_size)
dec_out = np.array_split(dec_out, n3d // batch_size)
# store all pose hypotheses in a list
pose_3d_mdm = [[], [], [], [], []]
for bidx in range(len(enc_in)):
# Dropout probability 0 (keep probability 1) for sampling
dp = 1.0
loss, _, out_all_components = model.step(sess,
enc_in[bidx],
dec_out[bidx],
dp,
isTraining=False)
# denormalize the input 2d pose, ground truth 3d pose as well as 3d pose hypotheses from mdm
out_all_components = np.reshape(
out_all_components,
[-1, model.HUMAN_3D_SIZE + 2, model.num_models])
out_mean = out_all_components[:, :model.HUMAN_3D_SIZE, :]
enc_in[bidx] = data_utils.unNormalizeData(
enc_in[bidx], data_mean_2d, data_std_2d,
dim_to_ignore_2d)
dec_out[bidx] = data_utils.unNormalizeData(
dec_out[bidx], data_mean_3d, data_std_3d,
dim_to_ignore_3d)
poses3d = np.zeros(
(out_mean.shape[0], 96, out_mean.shape[-1]))
for j in range(out_mean.shape[-1]):
poses3d[:, :, j] = data_utils.unNormalizeData(
out_mean[:, :, j], data_mean_3d, data_std_3d,
dim_to_ignore_3d)
# extract the 17 joints
dtu3d = np.hstack(
(np.arange(3), dim_to_use_3d
)) if not (FLAGS.predict_14) else dim_to_use_3d
dec_out_17 = dec_out[bidx][:, dtu3d]
pose_3d_17 = poses3d[:, dtu3d, :]
sqerr = (pose_3d_17 -
np.expand_dims(dec_out_17, axis=2))**2
dists = np.zeros(
(sqerr.shape[0], n_joints, sqerr.shape[2]))
for m in range(dists.shape[-1]):
dist_idx = 0
for k in np.arange(0, n_joints * 3, 3):
dists[:, dist_idx, m] = np.sqrt(
np.sum(sqerr[:, k:k + 3, m], axis=1))
dist_idx = dist_idx + 1
[
pose_3d_mdm[i].append(poses3d[:, :, i])
for i in range(poses3d.shape[-1])
]
# Put all the poses together
enc_in, dec_out = map(np.vstack, [enc_in, dec_out])
for i in range(poses3d.shape[-1]):
pose_3d_mdm[i] = np.vstack(pose_3d_mdm[i])
# Convert back to world coordinates
if FLAGS.camera_frame:
N_CAMERAS = 4
N_JOINTS_H36M = 32
# Add global position back
dec_out = dec_out + np.tile(test_root_positions[key3d],
[1, N_JOINTS_H36M])
for i in range(poses3d.shape[-1]):
pose_3d_mdm[i] = pose_3d_mdm[i] + np.tile(
test_root_positions[key3d], [1, N_JOINTS_H36M])
# Load the appropriate camera
subj, action, sname = key3d
cname = sname.split('.')[1] # <-- camera name
scams = {(subj, c + 1): rcams[(subj, c + 1)]
for c in range(N_CAMERAS)} # cams of this subject
scam_idx = [
scams[(subj, c + 1)][-1] for c in range(N_CAMERAS)
].index(cname) # index of camera used
the_cam = scams[(subj,
scam_idx + 1)] # <-- the camera used
R, T, f, c, k, p, name = the_cam
assert name == cname
def cam2world_centered(data_3d_camframe):
data_3d_worldframe = cameras.camera_to_world_frame(
data_3d_camframe.reshape((-1, 3)), R, T)
data_3d_worldframe = data_3d_worldframe.reshape(
(-1, N_JOINTS_H36M * 3))
# subtract root translation
return data_3d_worldframe - np.tile(
data_3d_worldframe[:, :3], (1, N_JOINTS_H36M))
# Apply inverse rotation and translation
dec_out = cam2world_centered(dec_out)
for i in range(poses3d.shape[-1]):
pose_3d_mdm[i] = cam2world_centered(pose_3d_mdm[i])
# sample some results to visualize
np.random.seed(42)
idx = np.random.permutation(enc_in.shape[0])
enc_in, dec_out = enc_in[idx, :], dec_out[idx, :]
for i in range(poses3d.shape[-1]):
pose_3d_mdm[i] = pose_3d_mdm[i][idx, :]
exidx = 1
nsamples = 20
for i in np.arange(nsamples):
fig = plt.figure(figsize=(20, 5))
subplot_idx = 1
gs1 = gridspec.GridSpec(1, 7) # 5 rows, 9 columns
gs1.update(wspace=-0.00,
hspace=0.05) # set the spacing between axes.
plt.axis('off')
# Plot 2d pose
ax1 = plt.subplot(gs1[subplot_idx - 1])
p2d = enc_in[exidx, :]
viz.show2Dpose(p2d, ax1)
ax1.invert_yaxis()
# Plot 3d gt
ax2 = plt.subplot(gs1[subplot_idx], projection='3d')
p3d = dec_out[exidx, :]
viz.show3Dpose(p3d, ax2)
# Plot 3d pose hypotheses
for i in range(poses3d.shape[-1]):
ax3 = plt.subplot(gs1[subplot_idx + i + 1],
projection='3d')
p3d = pose_3d_mdm[i][exidx]
viz.show3Dpose(p3d,
ax3,
lcolor="#9b59b6",
rcolor="#2ecc71")
# plt.show()
plt.savefig('{}/sample_{}_{}_{}_{}.png'.format(
path, subj, action, scam_idx, exidx))
plt.close(fig)
exidx = exidx + 1
def sample():
actions = data_utils.define_actions( FLAGS.action )
SUBJECT_IDS = [1,5,6,7,8,9,11]
rcams = load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14 )
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data( actions, FLAGS.data_dir, rcams )
print( "done reading and normalizing data." )
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto( device_count = device_count )) as sess:
print("Creating %d layers of %d units." % (FLAGS.num_layers, FLAGS.linear_size))
batch_size = 128
model = create_model(sess, actions, batch_size)
print("Model loaded")
for key2d in test_set_2d.keys():
(subj, b, fname) = key2d
print( "Subject: {}, action: {}, fname: {}".format(subj, b, fname) )
key3d = key2d if FLAGS.camera_frame else (subj, b, '{0}.h5'.format(fname.split('.')[0]))
key3d = (subj, b, fname[:-3]) if (fname.endswith('-sh')) and FLAGS.camera_frame else key3d
enc_in = test_set_2d[ key2d ]
n2d, _ = enc_in.shape
dec_out = test_set_3d[ key3d ]
n3d, _ = dec_out.shape
assert n2d == n3d
enc_in = np.array_split( enc_in, n2d // batch_size )
dec_out = np.array_split( dec_out, n3d // batch_size )
all_poses_3d = []
for bidx in range( len(enc_in) ):
dp = 1.0
_, _, poses3d = model.step(sess, enc_in[bidx], dec_out[bidx], dp, isTraining=False)
enc_in[bidx] = data_utils.unNormalizeData( enc_in[bidx], data_mean_2d, data_std_2d, dim_to_ignore_2d )
dec_out[bidx] = data_utils.unNormalizeData( dec_out[bidx], data_mean_3d, data_std_3d, dim_to_ignore_3d )
poses3d = data_utils.unNormalizeData( poses3d, data_mean_3d, data_std_3d, dim_to_ignore_3d )
all_poses_3d.append( poses3d )
enc_in, dec_out, poses3d = map( np.vstack, [enc_in, dec_out, all_poses_3d] )
if FLAGS.camera_frame:
N_CAMERAS = 4
N_JOINTS_H36M = 32
dec_out = dec_out + np.tile( test_root_positions[ key3d ], [1,N_JOINTS_H36M] )
subj, _, sname = key3d
cname = sname.split('.')[1]
scams = {(subj,c+1): rcams[(subj,c+1)] for c in range(N_CAMERAS)}
scam_idx = [scams[(subj,c+1)][-1] for c in range(N_CAMERAS)].index( cname )
the_cam = scams[(subj, scam_idx+1)]
R, T, f, c, k, p, name = the_cam
assert name == cname
def cam2world_centered(data_3d_camframe):
data_3d_worldframe = camera_to_world_frame(data_3d_camframe.reshape((-1, 3)), R, T)
data_3d_worldframe = data_3d_worldframe.reshape((-1, N_JOINTS_H36M*3))
return data_3d_worldframe - np.tile( data_3d_worldframe[:,:3], (1,N_JOINTS_H36M) )
dec_out = cam2world_centered(dec_out)
poses3d = cam2world_centered(poses3d)
enc_in, dec_out, poses3d = map( np.vstack, [enc_in, dec_out, poses3d] )
idx = np.random.permutation( enc_in.shape[0] )
enc_in, dec_out, poses3d = enc_in[idx, :], dec_out[idx, :], poses3d[idx, :]
import matplotlib.gridspec as gridspec
fig = plt.figure( figsize=(19.2, 10.8) )
gs1 = gridspec.GridSpec(5, 9)
gs1.update(wspace=-0.00, hspace=0.05)
plt.axis('off')
subplot_idx, exidx = 1, 1
nsamples = 15
for i in np.arange( nsamples ):
ax1 = plt.subplot(gs1[subplot_idx-1])
p2d = enc_in[exidx,:]
viz.show2Dpose( p2d, ax1 )
ax1.invert_yaxis()
ax2 = plt.subplot(gs1[subplot_idx], projection='3d')
p3d = dec_out[exidx,:]
viz.show3Dpose( p3d, ax2 )
ax3 = plt.subplot(gs1[subplot_idx+1], projection='3d')
p3d = poses3d[exidx,:]
viz.show3Dpose( p3d, ax3, lcolor="#9b59b6", rcolor="#2ecc71" )
exidx = exidx + 1
subplot_idx = subplot_idx + 3
plt.show()
def mpi():
actions = data_utils.define_actions(FLAGS.action)
number_of_actions = len(actions)
# Load camera parameters
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir + 'train_images.txt',
FLAGS.data_dir + 'valid_images.txt', FLAGS.data_dir + 'train.h5',
FLAGS.data_dir + 'valid.h5', FLAGS.camera_frame, rcams,
FLAGS.predict_14)
# Read stacked hourglass 2D predictions if use_sh, otherwise use groundtruth 2D projections
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(
actions, FLAGS.data_dir + 'train_images.txt',
FLAGS.data_dir + 'valid_images.txt', FLAGS.data_dir + 'train.h5',
FLAGS.data_dir + 'valid.h5')
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data(
actions, FLAGS.data_dir + 'train_images.txt',
FLAGS.data_dir + 'valid_images.txt', FLAGS.data_dir + 'train.h5',
FLAGS.data_dir + 'valid.h5', rcams)
print("done reading and normalizing data.")
# Load MPI data for testing:
mpi_test3d, mpi_mean3d, mpi_std3d, mpi_test2d, mpi_mean2d, mpi_std2d = data_utils.read_mpi(
'/mnt/lustre/xingyifei/test_3dhp/annotTest.h5', True, data_mean_2d,
data_mean_3d)
# Avoid using the GPU if requested
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(device_count=device_count,
allow_soft_placement=True)) as sess:
# === Create the model ===
print("Creating %d bi-layers of %d units." %
(FLAGS.num_layers, FLAGS.linear_size))
model = create_model(sess, actions, FLAGS.batch_size)
model.train_writer.add_graph(sess.graph)
print("Model created")
#=== This is the training loop ===
step_time, loss, val_loss = 0.0, 0.0, 0.0
current_step = 0 if FLAGS.load <= 0 else FLAGS.load + 1
previous_losses = []
step_time, loss = 0, 0
current_epoch = 0
log_every_n_batches = 100
# === Testing after this epoch ===
isTraining = False
n_joints = 16 if not (FLAGS.predict_14) else 14
#Process inputs to batches
n = len(mpi_test3d)
n_extra = n % model.batch_size
if n_extra > 0: # Otherwise examples are already a multiple of batch size
encoder_inputs = mpi_test2d[:-n_extra, :, :].reshape(-1, 32)
decoder_outputs = mpi_test3d[:-n_extra, :, :].reshape(-1, 48)
n_batches = n // model.batch_size
encoder_inputs = np.split(encoder_inputs, n_batches)
decoder_outputs = np.split(decoder_outputs, n_batches)
nbatches = len(encoder_inputs)
# Loop through test examples
all_dists, start_time, loss = [], time.time(), 0.
log_every_n_batches = 100
for i in range(nbatches):
if current_epoch > 0 and (i + 1) % log_every_n_batches == 0:
print("Working on test epoch {0}, batch {1} / {2}".format(
current_epoch, i + 1, nbatches))
enc_in, dec_out = encoder_inputs[i], decoder_outputs[i]
dp = 1.0 # dropout keep probability is always 1 at test time
step_loss, loss_summary, poses3d = model.step(sess,
enc_in,
dec_out,
dp,
isTraining=False)
loss += step_loss
# denormalize
enc_in = (enc_in.reshape(-1, 16, 2) * mpi_std2d +
mpi_mean2d).reshape(-1, 32)
dec_out = (dec_out.reshape(-1, 16, 3) * mpi_std3d +
mpi_mean3d).reshape(-1, 48)
poses3d = (poses3d.reshape(-1, 16, 3) * mpi_std3d +
mpi_mean3d).reshape(-1, 48)
assert dec_out.shape[0] == FLAGS.batch_size
assert poses3d.shape[0] == FLAGS.batch_size
if FLAGS.procrustes:
# Apply per-frame procrustes alignment if asked to do so
for j in range(FLAGS.batch_size):
gt = np.reshape(dec_out[j, :], [-1, 3])
out = np.reshape(poses3d[j, :], [-1, 3])
_, Z, T, b, c = procrustes.compute_similarity_transform(
gt, out, compute_optimal_scale=True)
out = (b * out.dot(T)) + c
poses3d[j, :] = np.reshape(out, [-1, 16 * 3]) if not (
FLAGS.predict_14) else np.reshape(out, [-1, 14 * 3])
# Compute Euclidean distance error per joint
sqerr = (
poses3d - dec_out
)**2 # Squared error between prediction and expected output
dists = np.zeros((sqerr.shape[0],
n_joints)) # Array with L2 error per joint in mm
dist_idx = 0
for k in np.arange(0, n_joints * 3, 3):
# Sum across X,Y, and Z dimenstions to obtain L2 distance
dists[:, dist_idx] = np.sqrt(np.sum(sqerr[:, k:k + 3], axis=1))
dist_idx = dist_idx + 1
all_dists.append(dists)
assert sqerr.shape[0] == FLAGS.batch_size
step_time = (time.time() - start_time) / nbatches
loss = loss / nbatches
all_dists = np.vstack(all_dists)
PCK_150 = all_dists < 150
from pdb import set_trace as st
st()
#AUC Metric
non_zero = np.count_nonzero(PCK_150, axis=1) * 1.
zeros = np.ones(non_zero.shape) * len(PCK_150[0]) - non_zero
TP = non_zero / len(PCK_150[0])
FP = zeros / len(PCK_150[0])
AUC = (1 + TP - FP) / 2.
AUC = np.mean(AUC)
#PCK@150
PCK = np.mean(np.mean(PCK_150, axis=1))
# Error per joint and total for all passed batches
joint_err = np.mean(all_dists, axis=0)
total_err = np.mean(all_dists)
print("=============================\n"
"Step-time (ms): %.4f\n"
"Val loss avg: %.4f\n"
"Val error avg (mm): %.2f\n"
"=============================" %
(1000 * step_time, loss, total_err))
print("PCK ERROR: " + str(PCK))
print("AUC ERROR: " + str(AUC))
for i in range(n_joints):
# 6 spaces, right-aligned, 5 decimal places
print("Error in joint {0:02d} (mm): {1:>5.2f}".format(
i + 1, joint_err[i]))
print("=============================")
def test():
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11]
actions = data_utils.define_actions(FLAGS.action)
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
"""
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils_org.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14)
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils_org.read_2d_predictions(actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils_org.create_2d_data( actions, FLAGS.data_dir, rcams )
print( "done reading and normalizing data." )
"""
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions, augmented3d, train_set_3d_for_noisy, data_mean_3d_test, data_std_3d_test = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14,
FLAGS.augment_all, FLAGS.augment_rot, FLAGS.augment_flip,
FLAGS.augment_trans, FLAGS.add_noise, FLAGS.add_kinematics)
# Read stacked hourglass 2D predictions if use_sh, otherwise use groundtruth 2D projections
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(
actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d, train_set_2d_for_noisy, data_mean_2d_test, data_std_2d_test = data_utils.create_2d_data(
actions, FLAGS.data_dir, rcams, augmented3d)
print("done reading and normalizing data.")
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)
with tf.Session(config=tf.ConfigProto(device_count=device_count,
gpu_options=gpu_options)) as sess:
# === Create the model ===
print("Creating %d layers of %d units." %
(FLAGS.num_layers, FLAGS.linear_size))
batch_size = FLAGS.batch_size
model = create_model(sess, actions, batch_size)
print("Model loaded")
if FLAGS.evaluateActionWise:
print("{0:=^12} {1:=^6}".format("Action",
"mm")) # line of 30 equal signs
cum_err = 0
for action in actions:
print("{0:<12} ".format(action), end="")
# Get 2d and 3d testing data for this action
action_test_set_2d = get_action_subset(test_set_2d, action)
action_test_set_3d = get_action_subset(test_set_3d, action)
encoder_inputs, decoder_outputs = model.get_all_batches(
action_test_set_2d,
action_test_set_3d,
FLAGS.camera_frame,
training=False)
act_err, _, step_time, loss = evaluate_batches(
sess, model, data_mean_3d_test, data_std_3d_test,
dim_to_use_3d, dim_to_ignore_3d, data_mean_2d_test,
data_std_2d_test, dim_to_use_2d, dim_to_ignore_2d, 0,
encoder_inputs, decoder_outputs)
cum_err = cum_err + act_err
print("{0:>6.2f}".format(act_err))
summaries = sess.run(
model.err_mm_summary,
{model.err_mm: float(cum_err / float(len(actions)))})
model.test_writer.add_summary(summaries, current_step)
print("{0:<12} {1:>6.2f}".format("Average",
cum_err / float(len(actions))))
print("{0:=^19}".format(''))
else:
n_joints = 17 if not (FLAGS.predict_14) else 14
encoder_inputs, decoder_outputs = model.get_all_batches(
test_set_2d, test_set_3d, FLAGS.camera_frame, training=False)
total_err, joint_err, step_time, loss = evaluate_batches(
sess, model, data_mean_3d_test, data_std_3d_test,
dim_to_use_3d, dim_to_ignore_3d, data_mean_2d_test,
data_std_2d_test, dim_to_use_2d, dim_to_ignore_2d, 0,
encoder_inputs, decoder_outputs, FLAGS.epochs)
print("=============================\n"
"Step-time (ms): %.4f\n"
"Val loss avg: %.4f\n"
"Val error avg (mm): %.2f\n"
"=============================" %
(1000 * step_time, loss, total_err))
for i in range(n_joints):
# 6 spaces, right-aligned, 5 decimal places
print("Error in joint {0:02d} (mm): {1:>5.2f}".format(
i + 1, joint_err[i]))
print("=============================")
def train():
actions = data_utils.define_actions( FLAGS.action )
number_of_actions = len( actions )
SUBJECT_IDS = [1,5,6,7,8,9,11]
rcams = load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14 )
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data( actions, FLAGS.data_dir, rcams )
print( "done reading and normalizing data." )
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(
device_count=device_count,
allow_soft_placement=True )) as sess:
print("Creating %d bi-layers of %d units." % (FLAGS.num_layers, FLAGS.linear_size))
model = create_model( sess, actions, FLAGS.batch_size )
model.train_writer.add_graph( sess.graph )
print("Model created")
step_time, loss, val_loss = 0.0, 0.0, 0.0
current_step = 0 if FLAGS.load <= 0 else FLAGS.load + 1
previous_losses = []
step_time, loss = 0, 0
current_epoch = 0
log_every_n_batches = 100
for _ in xrange( FLAGS.epochs ):
current_epoch = current_epoch + 1
encoder_inputs, decoder_outputs = model.get_all_batches( train_set_2d, train_set_3d, FLAGS.camera_frame, training=True )
nbatches = len( encoder_inputs )
print("There are {0} train batches".format( nbatches ))
start_time, loss = time.time(), 0.
for i in range( nbatches ):
if (i+1) % log_every_n_batches == 0:
print("Working on epoch {0}, batch {1} / {2}... ".format( current_epoch, i+1, nbatches), end="" )
enc_in, dec_out = encoder_inputs[i], decoder_outputs[i]
step_loss, loss_summary, lr_summary, _ = model.step( sess, enc_in, dec_out, FLAGS.dropout, isTraining=True )
if (i+1) % log_every_n_batches == 0:
model.train_writer.add_summary( loss_summary, current_step )
model.train_writer.add_summary( lr_summary, current_step )
step_time = (time.time() - start_time)
start_time = time.time()
print("done in {0:.2f} ms".format( 1000*step_time / log_every_n_batches ) )
loss += step_loss
current_step += 1
loss = loss / nbatches
print("=============================\n"
"Global step: %d\n"
"Learning rate: %.2e\n"
"Train loss avg: %.4f\n"
"=============================" % (model.global_step.eval(),
model.learning_rate.eval(), loss) )
isTraining = False
if FLAGS.evaluateActionWise:
print("{0:=^12} {1:=^6}".format("Action", "mm"))
cum_err = 0
for action in actions:
print("{0:<12} ".format(action), end="")
action_test_set_2d = get_action_subset( test_set_2d, action )
action_test_set_3d = get_action_subset( test_set_3d, action )
encoder_inputs, decoder_outputs = model.get_all_batches( action_test_set_2d, action_test_set_3d, FLAGS.camera_frame, training=False)
act_err, _, step_time, loss = evaluate_batches( sess, model,
data_mean_3d, data_std_3d, dim_to_use_3d, dim_to_ignore_3d,
data_mean_2d, data_std_2d, dim_to_use_2d, dim_to_ignore_2d,
current_step, encoder_inputs, decoder_outputs )
cum_err = cum_err + act_err
print("{0:>6.2f}".format(act_err))
summaries = sess.run( model.err_mm_summary, {model.err_mm: float(cum_err/float(len(actions)))} )
model.test_writer.add_summary( summaries, current_step )
print("{0:<12} {1:>6.2f}".format("Average", cum_err/float(len(actions) )))
print("{0:=^19}".format(''))
else:
n_joints = 17 if not(FLAGS.predict_14) else 14
encoder_inputs, decoder_outputs = model.get_all_batches( test_set_2d, test_set_3d, FLAGS.camera_frame, training=False)
total_err, joint_err, step_time, loss = evaluate_batches( sess, model,
data_mean_3d, data_std_3d, dim_to_use_3d, dim_to_ignore_3d,
data_mean_2d, data_std_2d, dim_to_use_2d, dim_to_ignore_2d,
current_step, encoder_inputs, decoder_outputs, current_epoch )
print("=============================\n"
"Step-time (ms): %.4f\n"
"Val loss avg: %.4f\n"
"Val error avg (mm): %.2f\n"
"=============================" % ( 1000*step_time, loss, total_err ))
for i in range(n_joints):
print("Error in joint {0:02d} (mm): {1:>5.2f}".format(i+1, joint_err[i]))
print("=============================")
summaries = sess.run( model.err_mm_summary, {model.err_mm: total_err} )
model.test_writer.add_summary( summaries, current_step )
print( "Saving the model... ", end="" )
start_time = time.time()
model.saver.save(sess, os.path.join(train_dir, 'checkpoint'), global_step=current_step )
print( "done in {0:.2f} ms".format(1000*(time.time() - start_time)) )
step_time, loss = 0, 0
sys.stdout.flush()
def sample():
"""Get samples from a model and visualize them"""
actions = data_utils.define_actions( FLAGS.action )
# Load camera parameters
SUBJECT_IDS = [1,5,6,7,8,9,11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14 )
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data( actions, FLAGS.data_dir, rcams )
print( "done reading and normalizing data." )
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto( device_count = device_count )) as sess:
# === Create the model ===
print("Creating %d layers of %d units." % (FLAGS.num_layers, FLAGS.linear_size))
batch_size = 128
model = create_model(sess, actions, batch_size)
print("Model loaded")
for key2d in test_set_2d.keys():
(subj, b, fname) = key2d
print( "Subject: {}, action: {}, fname: {}".format(subj, b, fname) )
# keys should be the same if 3d is in camera coordinates
key3d = key2d if FLAGS.camera_frame else (subj, b, '{0}.h5'.format(fname.split('.')[0]))
key3d = (subj, b, fname[:-3]) if (fname.endswith('-sh')) and FLAGS.camera_frame else key3d
enc_in = test_set_2d[ key2d ]
n2d, _ = enc_in.shape
dec_out = test_set_3d[ key3d ]
n3d, _ = dec_out.shape
assert n2d == n3d
# Split into about-same-size batches
enc_in = np.array_split( enc_in, n2d // batch_size )
dec_out = np.array_split( dec_out, n3d // batch_size )
all_poses_3d = []
for bidx in range( len(enc_in) ):
# Dropout probability 0 (keep probability 1) for sampling
dp = 1.0
_, _, poses3d = model.step(sess, enc_in[bidx], dec_out[bidx], dp, isTraining=False)
# denormalize
enc_in[bidx] = data_utils.unNormalizeData( enc_in[bidx], data_mean_2d, data_std_2d, dim_to_ignore_2d )
dec_out[bidx] = data_utils.unNormalizeData( dec_out[bidx], data_mean_3d, data_std_3d, dim_to_ignore_3d )
poses3d = data_utils.unNormalizeData( poses3d, data_mean_3d, data_std_3d, dim_to_ignore_3d )
all_poses_3d.append( poses3d )
# Put all the poses together
enc_in, dec_out, poses3d = map( np.vstack, [enc_in, dec_out, all_poses_3d] )
# Convert back to world coordinates
if FLAGS.camera_frame:
N_CAMERAS = 4
N_JOINTS_H36M = 32
# Add global position back
dec_out = dec_out + np.tile( test_root_positions[ key3d ], [1,N_JOINTS_H36M] )
# Load the appropriate camera
subj, _, sname = key3d
cname = sname.split('.')[1] # <-- camera name
scams = {(subj,c+1): rcams[(subj,c+1)] for c in range(N_CAMERAS)} # cams of this subject
scam_idx = [scams[(subj,c+1)][-1] for c in range(N_CAMERAS)].index( cname ) # index of camera used
the_cam = scams[(subj, scam_idx+1)] # <-- the camera used
R, T, f, c, k, p, name = the_cam
assert name == cname
def cam2world_centered(data_3d_camframe):
data_3d_worldframe = cameras.camera_to_world_frame(data_3d_camframe.reshape((-1, 3)), R, T)
data_3d_worldframe = data_3d_worldframe.reshape((-1, N_JOINTS_H36M*3))
# subtract root translation
return data_3d_worldframe - np.tile( data_3d_worldframe[:,:3], (1,N_JOINTS_H36M) )
# Apply inverse rotation and translation
dec_out = cam2world_centered(dec_out)
poses3d = cam2world_centered(poses3d)
# Grab a random batch to visualize
enc_in, dec_out, poses3d = map( np.vstack, [enc_in, dec_out, poses3d] )
idx = np.random.permutation( enc_in.shape[0] )
enc_in, dec_out, poses3d = enc_in[idx, :], dec_out[idx, :], poses3d[idx, :]
# Visualize random samples
import matplotlib.gridspec as gridspec
# 1080p = 1,920 x 1,080
fig = plt.figure( figsize=(19.2, 10.8) )
gs1 = gridspec.GridSpec(5, 9) # 5 rows, 9 columns
gs1.update(wspace=-0.00, hspace=0.05) # set the spacing between axes.
plt.axis('off')
subplot_idx, exidx = 1, 1
nsamples = 15
for i in np.arange( nsamples ):
# Plot 2d pose
ax1 = plt.subplot(gs1[subplot_idx-1])
p2d = enc_in[exidx,:]
viz.show2Dpose( p2d, ax1 )
ax1.invert_yaxis()
# Plot 3d gt
ax2 = plt.subplot(gs1[subplot_idx], projection='3d')
p3d = dec_out[exidx,:]
viz.show3Dpose( p3d, ax2 )
# Plot 3d predictions
ax3 = plt.subplot(gs1[subplot_idx+1], projection='3d')
p3d = poses3d[exidx,:]
viz.show3Dpose( p3d, ax3, lcolor="#9b59b6", rcolor="#2ecc71" )
exidx = exidx + 1
subplot_idx = subplot_idx + 3
plt.show()
def train():
"""Train a linear model for 3d pose estimation"""
actions = data_utils.define_actions( FLAGS.action )
number_of_actions = len( actions )
# Load camera parameters
SUBJECT_IDS = [1,5,6,7,8,9,11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14 )
# Read stacked hourglass 2D predictions if use_sh, otherwise use groundtruth 2D projections
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data( actions, FLAGS.data_dir, rcams )
print( "done reading and normalizing data." )
# Avoid using the GPU if requested
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(
device_count=device_count,
allow_soft_placement=True )) as sess:
# === Create the model ===
print("Creating %d bi-layers of %d units." % (FLAGS.num_layers, FLAGS.linear_size))
model = create_model( sess, actions, FLAGS.batch_size )
model.train_writer.add_graph( sess.graph )
print("Model created")
#=== This is the training loop ===
step_time, loss, val_loss = 0.0, 0.0, 0.0
current_step = 0 if FLAGS.load <= 0 else FLAGS.load + 1
previous_losses = []
step_time, loss = 0, 0
current_epoch = 0
log_every_n_batches = 100
for _ in xrange( FLAGS.epochs ):
current_epoch = current_epoch + 1
# === Load training batches for one epoch ===
encoder_inputs, decoder_outputs = model.get_all_batches( train_set_2d, train_set_3d, FLAGS.camera_frame, training=True )
nbatches = len( encoder_inputs )
print("There are {0} train batches".format( nbatches ))
start_time, loss = time.time(), 0.
# === Loop through all the training batches ===
for i in range( nbatches ):
if (i+1) % log_every_n_batches == 0:
# Print progress every log_every_n_batches batches
print("Working on epoch {0}, batch {1} / {2}... ".format( current_epoch, i+1, nbatches), end="" )
enc_in, dec_out = encoder_inputs[i], decoder_outputs[i]
step_loss, loss_summary, lr_summary, _ = model.step( sess, enc_in, dec_out, FLAGS.dropout, isTraining=True )
if (i+1) % log_every_n_batches == 0:
# Log and print progress every log_every_n_batches batches
model.train_writer.add_summary( loss_summary, current_step )
model.train_writer.add_summary( lr_summary, current_step )
step_time = (time.time() - start_time)
start_time = time.time()
print("done in {0:.2f} ms".format( 1000*step_time / log_every_n_batches ) )
loss += step_loss
current_step += 1
# === end looping through training batches ===
loss = loss / nbatches
print("=============================\n"
"Global step: %d\n"
"Learning rate: %.2e\n"
"Train loss avg: %.4f\n"
"=============================" % (model.global_step.eval(),
model.learning_rate.eval(), loss) )
# === End training for an epoch ===
# === Testing after this epoch ===
isTraining = False
if FLAGS.evaluateActionWise:
print("{0:=^12} {1:=^6}".format("Action", "mm")) # line of 30 equal signs
cum_err = 0
for action in actions:
print("{0:<12} ".format(action), end="")
# Get 2d and 3d testing data for this action
action_test_set_2d = get_action_subset( test_set_2d, action )
action_test_set_3d = get_action_subset( test_set_3d, action )
encoder_inputs, decoder_outputs = model.get_all_batches( action_test_set_2d, action_test_set_3d, FLAGS.camera_frame, training=False)
act_err, _, step_time, loss = evaluate_batches( sess, model,
data_mean_3d, data_std_3d, dim_to_use_3d, dim_to_ignore_3d,
data_mean_2d, data_std_2d, dim_to_use_2d, dim_to_ignore_2d,
current_step, encoder_inputs, decoder_outputs )
cum_err = cum_err + act_err
print("{0:>6.2f}".format(act_err))
summaries = sess.run( model.err_mm_summary, {model.err_mm: float(cum_err/float(len(actions)))} )
model.test_writer.add_summary( summaries, current_step )
print("{0:<12} {1:>6.2f}".format("Average", cum_err/float(len(actions) )))
print("{0:=^19}".format(''))
else:
n_joints = 17 if not(FLAGS.predict_14) else 14
encoder_inputs, decoder_outputs = model.get_all_batches( test_set_2d, test_set_3d, FLAGS.camera_frame, training=False)
total_err, joint_err, step_time, loss = evaluate_batches( sess, model,
data_mean_3d, data_std_3d, dim_to_use_3d, dim_to_ignore_3d,
data_mean_2d, data_std_2d, dim_to_use_2d, dim_to_ignore_2d,
current_step, encoder_inputs, decoder_outputs, current_epoch )
print("=============================\n"
"Step-time (ms): %.4f\n"
"Val loss avg: %.4f\n"
"Val error avg (mm): %.2f\n"
"=============================" % ( 1000*step_time, loss, total_err ))
for i in range(n_joints):
# 6 spaces, right-aligned, 5 decimal places
print("Error in joint {0:02d} (mm): {1:>5.2f}".format(i+1, joint_err[i]))
print("=============================")
# Log the error to tensorboard
summaries = sess.run( model.err_mm_summary, {model.err_mm: total_err} )
model.test_writer.add_summary( summaries, current_step )
# Save the model
print( "Saving the model... ", end="" )
start_time = time.time()
model.saver.save(sess, os.path.join(train_dir, 'checkpoint'), global_step=current_step )
print( "done in {0:.2f} ms".format(1000*(time.time() - start_time)) )
# Reset global time and loss
step_time, loss = 0, 0
sys.stdout.flush()
def train():
"""Train a linear model for 3d pose estimation"""
actions = data_utils.define_actions(
FLAGS.action) #returns a list of corresponding actions
number_of_actions = len(actions)
# Load camera parameters
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11] #1,5,6,7,8 for train, 9,11 for test
rcams = cameras.load_cameras(FLAGS.cameras_path,
SUBJECT_IDS) #得到了关于camera的参数
#print("相机的参数")
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14)
#print(data_mean_3d.shape[1],"平均值的大小")
#assert 1==2,"debug结束"
#train_set_3d是个dict，key为(S的编码，action，文件名)+pose（n,96)的大小
# Read stacked hourglass 2D predictions if use_sh, otherwise use groundtruth 2D projections
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(
actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data(
actions, FLAGS.data_dir, rcams)
print("done reading and normalizing data", test_set_2d.shape)
# Avoid using the GPU if requested
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(device_count=device_count,
allow_soft_placement=True)) as sess:
# === Create the model ===
print("Creating %d bi-layers of %d units." %
(FLAGS.num_layers,
FLAGS.linear_size)) #打印出来的是Creating 2 bi-layers of 1024 units.
model = create_model(sess, actions,
FLAGS.batch_size) #FLAGS的batch_size是64
model.train_writer.add_graph(sess.graph) #将图添加到tensorboard中
print("Model created")
#=== This is the training loop ===
step_time, loss, val_loss = 0.0, 0.0, 0.0
current_step = 0 if FLAGS.load <= 0 else FLAGS.load + 1 #之后需要载入checkpoint
previous_losses = []
step_time, loss = 0, 0
current_epoch = 0
log_every_n_batches = 100 #每100次打印一下
for _ in xrange(FLAGS.epochs): #与range类似但也有不同之处
current_epoch = current_epoch + 1
# === Load training batches for one epoch === 在这里输入和输出的还是32和48为的
encoder_inputs, decoder_outputs = model.get_all_batches(
train_set_2d, train_set_3d, FLAGS.camera_frame, training=True)
#按照batch的大小对输入和输出进行切片，切成[array(n1,32/48),array(n2,32/48),...,...]
nbatches = len(encoder_inputs)
print("There are {0} train batches".format(
nbatches)) #24371个branches
start_time, loss = time.time(), 0.
# === Loop through all the training batches ===
for i in range(nbatches):
if (i + 1) % log_every_n_batches == 0:
# Print progress every log_every_n_batches batches
print("Working on epoch {0}, batch {1} / {2}... ".format(
current_epoch, i + 1, nbatches),
end="")
enc_in, dec_out = encoder_inputs[i], decoder_outputs[i]
#print(enc_in.shape,dec_out.shape) (64,32)和(64,48)
#将输入值输入到model模型中，
step_loss, loss_summary, lr_summary, _ = model.step(
sess, enc_in, dec_out, FLAGS.dropout, isTraining=True)
if (i + 1) % log_every_n_batches == 0:
# Log and print progress every log_every_n_batches batches 每100次就将结果写到tensorboard中去
model.train_writer.add_summary(loss_summary, current_step)
model.train_writer.add_summary(lr_summary, current_step)
step_time = (time.time() - start_time)
start_time = time.time()
print("done in {0:.2f} ms".format(1000 * step_time /
log_every_n_batches))
loss += step_loss
current_step += 1
# === end looping through training batches ===
loss = loss / nbatches
print("=============================\n"
"Global step: %d\n"
"Learning rate: %.2e\n"
"Train loss avg: %.4f\n"
"=============================" %
(model.global_step.eval(), model.learning_rate.eval(), loss))
# === End training for an epoch ===
# === Testing after this epoch ===
isTraining = False
if FLAGS.evaluateActionWise:
print("{0:=^12} {1:=^6}".format(
"Action",
"mm")) # line of 30 equal signs 即为====Action====mm
cum_err = 0
for action in actions:
print("{0:<12} ".format(action), end="")
# Get 2d and 3d testing data for this action 并将他们按照batch的大小切割好（不需要训练的话，就不需要随机排列了
action_test_set_2d = get_action_subset(test_set_2d, action)
action_test_set_3d = get_action_subset(test_set_3d, action)
encoder_inputs, decoder_outputs = model.get_all_batches(
action_test_set_2d,
action_test_set_3d,
FLAGS.camera_frame,
training=False)
#evaluate_batches要好好看一下
act_err, _, step_time, loss = evaluate_batches(
sess, model, data_mean_3d, data_std_3d, dim_to_use_3d,
dim_to_ignore_3d, data_mean_2d, data_std_2d,
dim_to_use_2d, dim_to_ignore_2d, current_step,
encoder_inputs, decoder_outputs)
cum_err = cum_err + act_err
print("{0:>6.2f}".format(act_err))
summaries = sess.run(
model.err_mm_summary,
{model.err_mm: float(cum_err / float(len(actions)))})
model.test_writer.add_summary(summaries, current_step)
print("{0:<12} {1:>6.2f}".format("Average", cum_err /
float(len(actions))))
print("{0:=^19}".format(''))
else:
n_joints = 17 if not (FLAGS.predict_14) else 14
encoder_inputs, decoder_outputs = model.get_all_batches(
test_set_2d,
test_set_3d,
FLAGS.camera_frame,
training=False)
total_err, joint_err, step_time, loss = evaluate_batches(
sess, model, data_mean_3d, data_std_3d, dim_to_use_3d,
dim_to_ignore_3d, data_mean_2d, data_std_2d, dim_to_use_2d,
dim_to_ignore_2d, current_step, encoder_inputs,
decoder_outputs, current_epoch)
print("=============================\n"
"Step-time (ms): %.4f\n"
"Val loss avg: %.4f\n"
"Val error avg (mm): %.2f\n"
"=============================" %
(1000 * step_time, loss, total_err))
for i in range(n_joints):
# 6 spaces, right-aligned, 5 decimal places
print("Error in joint {0:02d} (mm): {1:>5.2f}".format(
i + 1, joint_err[i]))
print("=============================")
# Log the error to tensorboard
summaries = sess.run(model.err_mm_summary,
{model.err_mm: total_err})
model.test_writer.add_summary(summaries, current_step)
# Save the model
print("Saving the model... ", end="")
start_time = time.time()
model.saver.save(sess,
os.path.join(train_dir, 'checkpoint'),
global_step=current_step)
print("done in {0:.2f} ms".format(1000 *
(time.time() - start_time)))
# Reset global time and loss
step_time, loss = 0, 0
sys.stdout.flush()
def sample():
"""Get samples from a model and visualize them"""
actions = data_utils.define_actions( FLAGS.action )
# Load camera parameters
SUBJECT_IDS = [1,5,6,7,8,9,11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14 )
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data( actions, FLAGS.data_dir, rcams )
print( "done reading and normalizing data." )
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto( device_count = device_count )) as sess:
# === Create the model ===
print("Creating %d layers of %d units." % (FLAGS.num_layers, FLAGS.linear_size))
batch_size = 128
model = create_model(sess, actions, batch_size)
print("Model loaded")
for key2d in test_set_2d.keys():
(subj, b, fname) = key2d
print( "Subject: {}, action: {}, fname: {}".format(subj, b, fname) )
# keys should be the same if 3d is in camera coordinates
key3d = key2d if FLAGS.camera_frame else (subj, b, '{0}.h5'.format(fname.split('.')[0]))
key3d = (subj, b, fname[:-3]) if (fname.endswith('-sh')) and FLAGS.camera_frame else key3d
enc_in = test_set_2d[ key2d ]
n2d, _ = enc_in.shape
dec_out = test_set_3d[ key3d ]
n3d, _ = dec_out.shape
assert n2d == n3d
# Split into about-same-size batches
enc_in = np.array_split( enc_in, n2d // batch_size )
dec_out = np.array_split( dec_out, n3d // batch_size )
all_poses_3d = []
for bidx in range( len(enc_in) ):
# Dropout probability 0 (keep probability 1) for sampling
dp = 1.0
_, _, poses3d = model.step(sess, enc_in[bidx], dec_out[bidx], dp, isTraining=False)
# denormalize
enc_in[bidx] = data_utils.unNormalizeData( enc_in[bidx], data_mean_2d, data_std_2d, dim_to_ignore_2d )
dec_out[bidx] = data_utils.unNormalizeData( dec_out[bidx], data_mean_3d, data_std_3d, dim_to_ignore_3d )
poses3d = data_utils.unNormalizeData( poses3d, data_mean_3d, data_std_3d, dim_to_ignore_3d )
all_poses_3d.append( poses3d )
# Put all the poses together
enc_in, dec_out, poses3d = map( np.vstack, [enc_in, dec_out, all_poses_3d] )
# Convert back to world coordinates
if FLAGS.camera_frame:
N_CAMERAS = 4
N_JOINTS_H36M = 32
# Add global position back
dec_out = dec_out + np.tile( test_root_positions[ key3d ], [1,N_JOINTS_H36M] )
# Load the appropriate camera
subj, _, sname = key3d
cname = sname.split('.')[1] # <-- camera name
scams = {(subj,c+1): rcams[(subj,c+1)] for c in range(N_CAMERAS)} # cams of this subject
scam_idx = [scams[(subj,c+1)][-1] for c in range(N_CAMERAS)].index( cname ) # index of camera used
the_cam = scams[(subj, scam_idx+1)] # <-- the camera used
R, T, f, c, k, p, name = the_cam
assert name == cname
def cam2world_centered(data_3d_camframe):
data_3d_worldframe = cameras.camera_to_world_frame(data_3d_camframe.reshape((-1, 3)), R, T)
data_3d_worldframe = data_3d_worldframe.reshape((-1, N_JOINTS_H36M*3))
# subtract root translation
return data_3d_worldframe - np.tile( data_3d_worldframe[:,:3], (1,N_JOINTS_H36M) )
# Apply inverse rotation and translation
dec_out = cam2world_centered(dec_out)
poses3d = cam2world_centered(poses3d)
# Grab a random batch to visualize
enc_in, dec_out, poses3d = map( np.vstack, [enc_in, dec_out, poses3d] )
idx = np.random.permutation( enc_in.shape[0] )
enc_in, dec_out, poses3d = enc_in[idx, :], dec_out[idx, :], poses3d[idx, :]
# Visualize random samples
import matplotlib.gridspec as gridspec
# 1080p = 1,920 x 1,080
fig = plt.figure( figsize=(19.2, 10.8) )
gs1 = gridspec.GridSpec(5, 9) # 5 rows, 9 columns
gs1.update(wspace=-0.00, hspace=0.05) # set the spacing between axes.
plt.axis('off')
subplot_idx, exidx = 1, 1
nsamples = 15
for i in np.arange( nsamples ):
# Plot 2d pose
ax1 = plt.subplot(gs1[subplot_idx-1])
p2d = enc_in[exidx,:]
viz.show2Dpose( p2d, ax1 )
ax1.invert_yaxis()
# Plot 3d gt
ax2 = plt.subplot(gs1[subplot_idx], projection='3d')
p3d = dec_out[exidx,:]
viz.show3Dpose( p3d, ax2 )
# Plot 3d predictions
ax3 = plt.subplot(gs1[subplot_idx+1], projection='3d')
p3d = poses3d[exidx,:]
viz.show3Dpose( p3d, ax3, lcolor="#9b59b6", rcolor="#2ecc71" )
exidx = exidx + 1
subplot_idx = subplot_idx + 3
plt.show()
def predict(convert_to_world):
"""
Run the model and predict pose data
convert_to_world is a flag indicating whether to convert the data back to
world coordinates from the camera frame.
"""
actions = data_utils.define_actions(FLAGS.action)
# Load camera parameters
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14)
if FLAGS.use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(
actions, FLAGS.data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data(
actions, FLAGS.data_dir, rcams)
print("done reading and normalizing data.")
device_count = {"GPU": 0} if FLAGS.use_cpu else {"GPU": 1}
with tf.Session(config=tf.ConfigProto(device_count=device_count)) as sess:
# === Create the model ===
print("Creating %d layers of %d units." %
(FLAGS.num_layers, FLAGS.linear_size))
batch_size = 128
model = create_model(sess, actions, batch_size)
print("Model loaded")
for key2d in test_set_2d.keys():
(subj, b, fname) = key2d
print("Subject: {}, action: {}, fname: {}".format(subj, b, fname))
# keys should be the same if 3d is in camera coordinates
key3d = key2d if FLAGS.camera_frame else (subj, b, '{0}.h5'.format(
fname.split('.')[0]))
key3d = (subj, b, fname[:-3]) if (
fname.endswith('-sh')) and FLAGS.camera_frame else key3d
enc_in = test_set_2d[key2d]
n2d, _ = enc_in.shape
dec_out = test_set_3d[key3d]
n3d, _ = dec_out.shape
assert n2d == n3d
# Generate the loss pairs
loss_pairs = None
if model.num_loss_pairs:
num_pts = int(model.HUMAN_3D_SIZE / 3)
pairs = np.asarray([(i, j) for i in range(num_pts)
for j in range(num_pts) if i < j])
pair_idxs = [
np.random.choice(len(pairs),
model.num_loss_pairs,
replace=False) for _ in range(n3d)
]
loss_pairs = np.take(pairs, pair_idxs, axis=0)
loss_pairs = np.array_split(loss_pairs, n2d // batch_size)
# Split into about-same-size batches
enc_in = np.array_split(enc_in, n2d // batch_size)
dec_out = np.array_split(dec_out, n3d // batch_size)
all_poses_3d = []
# enc_in_modified = []
for bidx in range(len(enc_in)):
# Dropout probability 0 (keep probability 1) for sampling
dp = 1.0
if model.num_loss_pairs:
_, _, poses3d = model.step(sess,
enc_in[bidx],
dec_out[bidx],
dp,
loss_pairs=loss_pairs[bidx],
isTraining=False)
else:
_, _, poses3d = model.step(sess,
enc_in[bidx],
dec_out[bidx],
dp,
isTraining=False)
# poses3dnew = []
# for e in enc_in[bidx]:
# poses3dnew.append(np.insert(e, range(1, len(e)+1, 2), poses3d[1::3]))
# poses3d = poses3dnew
# print (bidx)
# print (len(enc_in[bidx]))
# print (enc_in[bidx])
# print (data_mean_2d)
# print (data_mean_3d)
# data_mean_2d_modified = np.delete(data_mean_3d, np.arange(2, data_mean_3d.size, 3))
# data_std_2d_modified = np.delete(data_std_3d, np.arange(2, data_std_3d.size, 3))
# denormalize
# enc_in_modified.append(data_utils.unNormalizeData( enc_in[bidx], data_mean_2d_modified, data_std_2d_modified, dim_to_ignore_2d ))
enc_in[bidx] = data_utils.unNormalizeData(
enc_in[bidx], data_mean_2d, data_std_2d, dim_to_ignore_2d)
dec_out[bidx] = data_utils.unNormalizeData(
dec_out[bidx], data_mean_3d, data_std_3d, dim_to_ignore_3d)
poses3d = data_utils.unNormalizeData(poses3d, data_mean_3d,
data_std_3d,
dim_to_ignore_3d)
all_poses_3d.append(poses3d)
# print (len(enc_in[bidx]))
# print (len(poses3d))
# print (len(enc_in[0]))
# print (len(poses3d[0]))
# Put all the poses together
# enc_in_modified = np.vstack(enc_in_modified)
enc_in, dec_out, poses3d = map(np.vstack,
[enc_in, dec_out, all_poses_3d])
# print (len(enc_in[0]))
# print (len(poses3d[0]))
# print (enc_in.shape)
# print (poses3d.shape)
# poses3dnew = []
# for p, e in zip(poses3d, enc_in_modified):
# poses3dnew.append(np.insert(e, range(1, len(e)+1, 2), p[1::3]))
# poses3d = np.array(poses3dnew)
if convert_to_world:
# Convert back to world coordinates
if FLAGS.camera_frame:
N_CAMERAS = 4
N_JOINTS_H36M = 32
# Add global position back
dec_out = dec_out + np.tile(test_root_positions[key3d],
[1, N_JOINTS_H36M])
# Load the appropriate camera
subj, _, sname = key3d
cname = sname.split('.')[1] # <-- camera name
scams = {(subj, c + 1): rcams[(subj, c + 1)]
for c in range(N_CAMERAS)} # cams of this subject
scam_idx = [
scams[(subj, c + 1)][-1] for c in range(N_CAMERAS)
].index(cname) # index of camera used
the_cam = scams[(subj,
scam_idx + 1)] # <-- the camera used
R, T, f, c, k, p, name = the_cam
assert name == cname
def cam2world_centered(data_3d_camframe):
data_3d_worldframe = cameras.camera_to_world_frame(
data_3d_camframe.reshape((-1, 3)), R, T)
data_3d_worldframe = data_3d_worldframe.reshape(
(-1, N_JOINTS_H36M * 3))
# subtract root translation
return data_3d_worldframe - np.tile(
data_3d_worldframe[:, :3], (1, N_JOINTS_H36M))
# Apply inverse rotation and translation
dec_out = cam2world_centered(dec_out)
poses3d = cam2world_centered(poses3d)
poses3dnew = dec_out.copy()
poses3dnew[:, 1::3] = poses3d[:, 1::3]
poses3d = poses3dnew
return enc_in, dec_out, poses3d
# Load data
data_dir = './data/h36m/'
camera_frame = True
predict_14 = False
# Load 3d data and load (or create) 2d projections
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, data_dir, camera_frame, rcams, predict_14 )
# Read stacked hourglass 2D predictions if use_sh, otherwise use groundtruth 2D projections
#use stacked hourgalss
use_sh = False
if use_sh:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(actions, data_dir)
else:
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data( actions, data_dir, rcams )
print( "done reading and normalizing data." )
stat_3d = {}
stat_3d['mean'] = data_mean_3d
stat_3d['std'] = data_std_3d
stat_3d['dim_use'] = dim_to_use_3d
# ============================
# Define Train/Test Methods
# ============================
def train(train_loader, model, criterion, optimizer,
lr_init=None, lr_now=None, glob_step=None, lr_decay=None, gamma=None,
max_norm=True):
losses = utils.AverageMeter()
def main(_):
smoothed = read_openpose_json_mydata(json_file=FLAGS.json)
plt.figure(2)
smooth_curves_plot = show_anim_curves(smoothed, plt)
# return
pngName = 'gif_output/smooth_plot.png'
smooth_curves_plot.savefig(pngName)
logger.info('writing gif_output/smooth_plot.png')
if FLAGS.interpolation:
logger.info("start interpolation")
framerange = len(smoothed.keys())
joint_rows = 36
array = np.concatenate(list(smoothed.values()))
array_reshaped = np.reshape(array, (framerange, joint_rows))
multiplier = FLAGS.multiplier
multiplier_inv = 1 / multiplier
out_array = np.array([])
for row in range(joint_rows):
x = []
for frame in range(framerange):
x.append(array_reshaped[frame, row])
frame = range(framerange)
frame_resampled = np.arange(0, framerange, multiplier)
spl = UnivariateSpline(frame, x, k=3)
# relative smooth factor based on jnt anim curve
min_x, max_x = min(x), max(x)
smooth_fac = max_x - min_x
smooth_resamp = 125
smooth_fac = smooth_fac * smooth_resamp
spl.set_smoothing_factor(float(smooth_fac))
xnew = spl(frame_resampled)
out_array = np.append(out_array, xnew)
logger.info(
"done interpolating. reshaping {0} frames, please wait!!".format(
framerange))
a = np.array([])
for frame in range(int(framerange * multiplier_inv)):
jnt_array = []
for jnt in range(joint_rows):
jnt_array.append(
out_array[jnt * int(framerange * multiplier_inv) + frame])
a = np.append(a, jnt_array)
a = np.reshape(a, (int(framerange * multiplier_inv), joint_rows))
out_array = a
interpolate_smoothed = {}
for frame in range(int(framerange * multiplier_inv)):
interpolate_smoothed[frame] = list(out_array[frame])
plt.figure(3)
smoothed = interpolate_smoothed
interpolate_curves_plot = show_anim_curves(smoothed, plt)
pngName = 'gif_output/interpolate_{0}.png'.format(smooth_resamp)
interpolate_curves_plot.savefig(pngName)
logger.info('writing gif_output/interpolate_plot.png')
enc_in = np.zeros((1, 64))
enc_in[0] = [0 for i in range(64)]
actions = data_utils.define_actions(FLAGS.action)
SUBJECT_IDS = [1, 5, 6, 7, 8, 9, 11]
rcams = cameras.load_cameras(FLAGS.cameras_path, SUBJECT_IDS)
# train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.read_2d_predictions(
# actions, FLAGS.data_dir)
train_set_2d, test_set_2d, data_mean_2d, data_std_2d, dim_to_ignore_2d, dim_to_use_2d = data_utils.create_2d_data(
actions, FLAGS.data_dir, rcams)
train_set_3d, test_set_3d, data_mean_3d, data_std_3d, dim_to_ignore_3d, dim_to_use_3d, train_root_positions, test_root_positions = data_utils.read_3d_data(
actions, FLAGS.data_dir, FLAGS.camera_frame, rcams, FLAGS.predict_14)
device_count = {"GPU": 1}
png_lib = []
before_pose = None
with tf.Session(config=tf.ConfigProto(device_count=device_count,
allow_soft_placement=True)) as sess:
# plt.figure(3)
batch_size = 128
model = create_model(sess, actions, batch_size)
iter_range = len(smoothed.keys())
export_units = {}
twod_export_units = {}
for n, (frame, xy) in enumerate(smoothed.items()):
logger.info("calc frame {0}/{1}".format(frame, iter_range))
# map list into np array
joints_array = np.zeros((1, 36))
joints_array[0] = [0 for i in range(36)]
for o in range(len(joints_array[0])):
# feed array with xy array
joints_array[0][o] = float(xy[o])
twod_export_units[frame] = {}
for abs_b, __n in enumerate(range(0, len(xy), 2)):
twod_export_units[frame][abs_b] = {
"translate": [xy[__n], xy[__n + 1]]
}
_data = joints_array[0]
# mapping all body parts or 3d-pose-baseline format
for i in range(len(order)):
for j in range(2):
# create encoder input
enc_in[0][order[i] * 2 + j] = _data[i * 2 + j]
for j in range(2):
# Hip
enc_in[0][0 * 2 + j] = (enc_in[0][1 * 2 + j] +
enc_in[0][6 * 2 + j]) / 2
# Neck/Nose
enc_in[0][14 * 2 + j] = (enc_in[0][15 * 2 + j] +
enc_in[0][12 * 2 + j]) / 2
# Thorax
enc_in[0][13 * 2 +
j] = 2 * enc_in[0][12 * 2 + j] - enc_in[0][14 * 2 +
j]
# set spine
spine_x = enc_in[0][24]
spine_y = enc_in[0][25]
enc_in = enc_in[:, dim_to_use_2d]
mu = data_mean_2d[dim_to_use_2d]
stddev = data_std_2d[dim_to_use_2d]
enc_in = np.divide((enc_in - mu), stddev)
dp = 1.0
# dec_out = np.zeros((1, 48))
# dec_out[0] = [0 for i in range(48)]
dec_out = np.zeros((1, 36))
dec_out[0] = [0 for i in range(36)]
_, _, poses3d = model.step(sess,
enc_in,
dec_out,
dp,
isTraining=False)
all_poses_3d = []
enc_in = data_utils.unNormalizeData(enc_in, data_mean_2d,
data_std_2d, dim_to_ignore_2d)
poses3d = data_utils.unNormalizeData(poses3d, data_mean_3d,
data_std_3d, dim_to_ignore_3d)
gs1 = gridspec.GridSpec(1, 1)
gs1.update(wspace=-0.00,
hspace=0.05) # set the spacing between axes.
plt.axis('off')
all_poses_3d.append(poses3d)
enc_in, poses3d = map(np.vstack, [enc_in, all_poses_3d])
subplot_idx, exidx = 1, 1
_max = 0
_min = 10000
for i in range(poses3d.shape[0]):
for j in range(32):
tmp = poses3d[i][j * 3 + 2]
poses3d[i][j * 3 + 2] = poses3d[i][j * 3 + 1]
poses3d[i][j * 3 + 1] = tmp
if poses3d[i][j * 3 + 2] > _max:
_max = poses3d[i][j * 3 + 2]
if poses3d[i][j * 3 + 2] < _min:
_min = poses3d[i][j * 3 + 2]
for i in range(poses3d.shape[0]):
for j in range(32):
poses3d[i][j * 3 + 2] = _max - poses3d[i][j * 3 + 2] + _min
poses3d[i][j * 3] += (spine_x - 630)
poses3d[i][j * 3 + 2] += (500 - spine_y)
# Plot 3d predictions
ax = plt.subplot(gs1[subplot_idx - 1], projection='3d')
ax.view_init(18, -70)
# if FLAGS.cache_on_fail:
# if np.min(poses3d) < -1000:
# poses3d = before_pose
p3d = poses3d
logger.info("frame score {0}".format(np.min(poses3d)))
x, y, z = [[] for _ in range(3)]
if not poses3d is None:
to_export = poses3d.tolist()[0]
else:
to_export = [0.0 for _ in range(96)]
logger.info("export {0}".format(to_export))
for o in range(0, len(to_export), 3):
x.append(to_export[o])
y.append(to_export[o + 1])
z.append(to_export[o + 2])
export_units[frame] = {}
for jnt_index, (_x, _y, _z) in enumerate(zip(x, y, z)):
export_units[frame][jnt_index] = {"translate": [_x, _y, _z]}
viz.show3Dpose(p3d, ax, lcolor="#9b59b6", rcolor="#2ecc71")
pngName = 'png/pose_frame_{0}.png'.format(str(frame).zfill(12))
plt.savefig(pngName)
if FLAGS.write_gif:
png_lib.append(imageio.imread(pngName))
# if FLAGS.cache_on_fail:
# before_pose = poses3d
if FLAGS.write_gif:
if FLAGS.interpolation:
# take every frame on gif_fps * multiplier_inv
png_lib = np.array([
png_lib[png_image]
for png_image in range(0, len(png_lib), int(multiplier_inv))
])
logger.info("creating Gif gif_output/animation.gif, please Wait!")
imageio.mimsave('gif_output/animation.gif', png_lib, fps=FLAGS.gif_fps)
_out_file = os.path.join(os.path.dirname(os.path.dirname(__file__)),
'maya/3d_data.json')
twod_out_file = os.path.join(os.path.dirname(os.path.dirname(__file__)),
'maya/2d_data.json')
with open(_out_file, 'w') as outfile:
logger.info("exported maya json to {0}".format(_out_file))
json.dump(export_units, outfile)
with open(twod_out_file, 'w') as outfile:
logger.info("exported maya json to {0}".format(twod_out_file))
json.dump(twod_export_units, outfile)
logger.info("Done!".format(pngName))
