def test_procedure(net, test_records):
# Load data
reader = SketchReader(tfrecord_list=test_records,
raw_size=[256, 256, 25],
shuffle=False,
num_threads=hyper_params['nbThreads'],
batch_size=1,
nb_epoch=1)
raw_input = reader.next_batch()
npr_lines, ds, _, fm, fm_inv, gt_normal, gt_depth, _, mask_cline_inv, mask_shape, mask_ds, mask_line, \
mask_line_inv, mask_2d, sel_mask, vdotn_scalar = net.cook_raw_inputs(raw_input)
# Network forward
logit_d, logit_n, logit_c, _ = net.load_baseline_net(
npr_lines,
ds,
mask_2d,
fm,
sel_mask,
vdotn_scalar,
hyper_params['rootFt'],
is_training=False)
# Test loss
test_loss, test_d_loss, test_n_loss, test_ds_loss, test_reg_loss, test_real_dloss, \
test_real_nloss, test_omega_loss, out_gt_normal, out_f_normal, out_gt_depth, out_f_depth, out_gt_ds, gt_lines, \
reg_mask, out_cf_map = loss(logit_d, logit_n, logit_c, gt_normal, gt_depth, mask_shape, mask_ds, mask_cline_inv,
ds, npr_lines, fm_inv)
return test_loss, test_d_loss, test_n_loss, test_ds_loss, test_reg_loss, test_real_dloss, \
test_real_nloss, test_omega_loss, out_gt_normal, out_f_normal, out_gt_depth, \
out_f_depth, out_gt_ds, gt_lines, reg_mask, out_cf_map
def validation_procedure(net, val_records, reg_weight):
# Load data
with tf.name_scope('eval_inputs') as _:
reader = SketchReader(tfrecord_list=val_records,
raw_size=[256, 256, 25],
shuffle=False,
num_threads=hyper_params['nbThreads'],
batch_size=hyper_params['batchSize'])
raw_input = reader.next_batch()
npr_lines, ds, _, fm, fm_inv, gt_normal, gt_depth, gt_field, mask_cline_inv, mask_shape, mask_ds, _, \
_, mask2d, selm, ndotv = net.cook_raw_inputs(raw_input)
# Network forward
logit_f, field_vars = net.load_field_net(npr_lines,
mask2d,
ds,
fm,
selm,
ndotv,
hyper_params['rootFt'],
is_training=False,
reuse=True)
logit_d, logit_n, logit_c, _ = net.load_GeomNet(npr_lines,
ds,
mask2d,
fm,
selm,
ndotv,
logit_f,
mask_cline_inv,
hyper_params['rootFt'],
is_training=False,
reuse=True)
# Validate loss
val_loss, val_d_loss, val_n_loss, val_ds_loss, val_reg_loss, val_real_dloss, val_real_nloss, \
val_omega_loss = loss(logit_d,
logit_n,
logit_c,
gt_normal,
gt_depth,
mask_shape,
mask_ds,
mask_cline_inv,
reg_weight,
fm_inv,
scope='test_loss')
# TensorBoard
proto_list = collect_vis_img_val(logit_d, logit_n, logit_c, logit_f,
npr_lines, gt_normal, gt_depth,
mask_shape, ds, gt_field, mask_cline_inv,
fm, fm_inv, selm, ndotv)
merged_val = tf.summary.merge(proto_list)
return merged_val, val_loss, val_d_loss, val_n_loss, val_ds_loss, val_reg_loss, val_real_dloss, \
val_real_nloss, val_omega_loss, field_vars
def validation_procedure(net, val_records, reg_weight):
# Load data
with tf.name_scope('eval_inputs') as _:
reader = SketchReader(tfrecord_list=val_records, raw_size=[256, 256, 25], shuffle=False,
num_threads=hyper_params['nbThreads'], batch_size=hyper_params['batchSize'])
raw_input = reader.next_batch()
npr_lines, ds, _, fm, fm_inv, gt_normal, gt_depth, _, mask_cline_inv, mask_shape, mask_ds, _, \
_, mask_2d, sel_mask, vdotn_scalar = net.cook_raw_inputs(raw_input)
# Network forward
logit_d, logit_n, logit_c, _ = net.load_baseline_net(nprLine_input,
ds_input,
mask2D_input,
fm_input,
selLineMask_input,
vdotnScalar_input,
hyper_params['rootFt'],
is_training=False,
reuse=True)
# Validate loss
val_loss, val_d_loss, val_n_loss, val_ds_loss, val_reg_loss, val_real_dloss, val_real_nloss, \
val_omega_loss = loss(logit_d,
logit_n,
logit_c,
gtNormal_input,
gtDepth_input,
maskShape_input,
maskDs_input,
clineInvMask_input,
reg_weight,
fmInv_input,
scope='test_loss')
# Tensorboard
proto_list = collect_vis_img_val(logit_d,
logit_n,
logit_c,
nprLine_input,
gtNormal_input,
gtDepth_input,
maskShape_input,
ds_input,
clineInvMask_input,
fm_input,
fmInv_input,
selLineMask_input,
vdotnScalar_input,
mask2D_input)
merged_val = tf.summary.merge(proto_list)
return merged_val, val_loss, val_d_loss, val_n_loss, val_ds_loss, val_reg_loss, val_real_dloss, \
val_real_nloss, val_omega_loss, [npr_lines, ds, fm, fm_inv, gt_normal, gt_depth, mask_cline_inv,
mask_shape, mask_ds, mask_2d, sel_mask, vdotn_scalar]
def test_procedure(net, test_records):
# Load data
reader = SketchReader(tfrecord_list=test_records, raw_size=[256, 256, 25],
shuffle=False, num_threads=hyper_params['nbThreads'],
batch_size=1, nb_epoch=1)
raw_input = reader.next_batch()
npr_lines, ds, _, fm, fm_inv, gt_normal, gt_depth, gt_field, mask_cline_inv, mask_shape, mask_ds, _, _, \
mask2d, selm, ndotv = net.cook_raw_inputs(raw_input)
# Network forward
logit_f, _ = net.load_field_net(nprLine_input,
mask2D_input,
ds_input,
fm_input,
selLineMask_input,
vdotnScalar_input,
hyper_params['rootFt'],
is_training=False)
logit_d, logit_n, logit_c, _ = net.load_GeomNet(nprLine_input,
ds_input,
mask2D_input,
fm_input,
selLineMask_input,
vdotnScalar_input,
logit_f,
clineInvMask_input,
hyper_params['rootFt'],
is_training=False)
# Test loss
test_loss, test_d_loss, test_n_loss, test_ds_loss, test_reg_loss, test_real_dloss, \
test_real_nloss, test_omega_loss, out_gt_normal, out_f_normal, out_gt_depth, out_f_depth, out_gt_ds, gt_lines, \
reg_mask, out_cf_map, test_gt_a, test_gt_b, test_f_a, test_f_b \
= loss(logit_d,
logit_n,
logit_c,
gtNormal_input,
gtDepth_input,
maskShape_input,
maskDs_input,
clineInvMask_input,
ds_input,
nprLine_input,
logit_f,
gtField_input,
fmInv_input)
return test_loss, test_d_loss, test_n_loss, test_ds_loss, test_reg_loss, test_real_dloss, \
test_real_nloss, test_omega_loss, out_gt_normal, out_f_normal, out_gt_depth, \
out_f_depth, out_gt_ds, gt_lines, reg_mask, out_cf_map, test_gt_a, test_gt_b, test_f_a, test_f_b, \
[npr_lines, ds, fm, fm_inv, gt_normal, gt_depth, gt_field,
mask_cline_inv, mask_shape, mask_ds, mask2d, selm, ndotv]
def validation_procedure(net, val_records, reg_weight):
# Load data
with tf.name_scope('eval_inputs') as _:
reader = SketchReader(tfrecord_list=val_records, raw_size=[256, 256, 25], shuffle=False,
num_threads=hyper_params['nbThreads'], batch_size=hyper_params['batchSize'])
raw_input = reader.next_batch()
npr_lines, ds, _, fm, fm_inv, gt_normal, gt_depth, _, mask_cline_inv, mask_shape, mask_ds, _, _, \
mask_2d, sel_mask, vdotn_scalar = net.cook_raw_inputs(raw_input)
# Network forward
logit_d, logit_n, _ = net.load_dn_naive_net(npr_lines,
ds,
mask_2d,
fm,
sel_mask,
vdotn_scalar,
hyper_params['rootFt'],
is_training=False,
reuse=True)
# Validate loss
val_loss, val_d_loss, val_n_loss, val_real_dloss, val_real_nloss, val_r_loss, val_ds_loss \
= loss(logit_d,
logit_n,
gt_normal,
gt_depth,
mask_shape,
mask_ds,
mask_cline_inv,
reg_weight,
fm_inv,
mode=hyper_params['lossId'],
scope='test_loss')
# TensorBoard
proto_list = collect_vis_img_val(logit_d,
logit_n,
npr_lines,
gt_normal,
gt_depth,
mask_shape,
ds,
mask_cline_inv,
fm,
fm_inv,
sel_mask,
vdotn_scalar,
mask_2d)
merged_val = tf.summary.merge(proto_list)
return merged_val, val_loss, val_d_loss, val_n_loss, val_real_dloss, val_real_nloss, val_r_loss, val_ds_loss
def validation_procedure(net, val_records):
# Load data
with tf.name_scope('eval_inputs') as _:
reader = SketchReader(tfrecord_list=val_records,
raw_size=[256, 256, 25],
shuffle=False,
num_threads=hyper_params['nbThreads'],
batch_size=hyper_params['batchSize'])
raw_input = reader.next_batch()
npr_lines, ds, _, fm, fm_inv, _, _, gt_field, cline_mask_inv, mask_shape, _, _, _, mask2d, selm, vdotn \
= net.cook_raw_inputs(raw_input)
# Network forward
logit_f, _ = net.load_field_net(nprLine_input,
mask2D_input,
ds_input,
fm_input,
selLineMask_input,
vdotnScalar_input,
hyper_params['rootFt'],
is_training=False,
reuse=True)
# Validate loss
val_loss, val_data_loss, val_smooth_loss = loss(logit_f,
gtField_input,
maskShape_input,
clineInvMask_input,
fmInv_input,
scope='test_loss')
# TensorBoard
proto_list = collect_vis_img_val(logit_f, nprLine_input, gtField_input,
maskShape_input, clineInvMask_input,
ds_input, fm_input, selLineMask_input,
vdotnScalar_input, fmInv_input)
merged_val = tf.summary.merge(proto_list)
return merged_val, val_loss, val_data_loss, val_smooth_loss, \
[npr_lines, ds, fm, fm_inv, gt_field, cline_mask_inv, mask_shape, mask2d, selm, vdotn]
def test_procedure(net, test_records):
# Load data
reader = SketchReader(tfrecord_list=test_records,
raw_size=[256, 256, 25],
shuffle=False,
num_threads=hyper_params['nbThreads'],
batch_size=1,
nb_epoch=1)
raw_input = reader.next_batch()
npr_lines, ds, _, fm, _, gt_normal, gt_depth, _, mask_cline_inv, mask_shape, mask_ds, _, _, mask_2d, \
sel_m, vdotn = net.cook_raw_inputs(raw_input)
# Network forward
logit_d, logit_n, _ = net.load_dn_naive_net(npr_lines,
ds,
mask_2d,
fm,
sel_m,
vdotn,
hyper_params['rootFt'],
is_training=False)
# Test loss
test_loss, test_d_loss, test_n_loss, test_real_dloss, test_real_nloss, out_gt_normal, out_f_normal, out_gt_depth, \
out_f_depth, out_gt_ds, gt_lines = loss(logit_d,
logit_n,
gt_normal,
gt_depth,
mask_shape,
mask_ds,
mask_cline_inv,
ds,
npr_lines,
mode=hyper_params['lossId'])
return test_loss, test_d_loss, test_n_loss, test_real_dloss, test_real_nloss, out_gt_normal, out_f_normal, \
out_gt_depth, out_f_depth, out_gt_ds, gt_lines
def train_procedure(net, train_records):
nb_gpus = hyper_params['nb_gpus']
# Load data
with tf.name_scope('train_inputs') as _:
bSize = hyper_params['batchSize'] * nb_gpus
nbThreads = hyper_params['nbThreads'] * nb_gpus
reader = SketchReader(tfrecord_list=train_records,
raw_size=[256, 256, 25],
shuffle=True,
num_threads=nbThreads,
batch_size=bSize)
raw_input = reader.next_batch()
npr_lines, ds, _, fm, fm_inv, _, _, gt_field, cline_mask_inv, mask_shape, _, _, _, mask_2d, selm, vdotn \
= net.cook_raw_inputs(raw_input)
# initialize optimizer
opt = tf.train.AdamOptimizer()
# split data
with tf.name_scope('divide_data'):
gpu_npr_lines = tf.split(nprLine_input, nb_gpus, axis=0)
gpu_mask2d = tf.split(mask2D_input, nb_gpus, axis=0)
gpu_ds = tf.split(ds_input, nb_gpus, axis=0)
gpu_fm = tf.split(fm_input, nb_gpus, axis=0)
gpu_fm_inv = tf.split(fmInv_input, nb_gpus, axis=0)
gpu_selm = tf.split(selLineMask_input, nb_gpus, axis=0)
gpu_vdotn = tf.split(vdotnScalar_input, nb_gpus, axis=0)
gpu_gt_field = tf.split(gtField_input, nb_gpus, axis=0)
gpu_mask_shape = tf.split(maskShape_input, nb_gpus, axis=0)
gpu_mask_cline = tf.split(clineInvMask_input, nb_gpus, axis=0)
tower_grads = []
tower_loss_collected = []
tower_total_losses = []
tower_data_losses = []
tower_smooth_losses = []
# TensorBoard: images
gpu0_npr_lines_imgs = None
gpu0_logit_f_imgs = None
gpu0_gt_f_imgs = None
gpu0_shape_mask_imgs = None
gpu0_shape_cline_imgs = None
gpu0_ds_imgs = None
gpu0_fm_imgs = None
gpu0_fm_inv_imgs = None
gpu0_selm_imgs = None
gpu0_vdotn_imgs = None
with tf.variable_scope(tf.get_variable_scope()):
for gpu_id in range(nb_gpus):
with tf.device('/gpu:%d' % gpu_id):
with tf.name_scope('tower_%s' % gpu_id) as _:
# Network forward
logit_f, _ = net.load_field_net(gpu_npr_lines[gpu_id],
gpu_mask2d[gpu_id],
gpu_ds[gpu_id],
gpu_fm[gpu_id],
gpu_selm[gpu_id],
gpu_vdotn[gpu_id],
hyper_params['rootFt'],
is_training=True)
# Training loss
train_loss, train_data_loss, train_smooth_loss = loss(
logit_f,
gpu_gt_field[gpu_id],
gpu_mask_shape[gpu_id],
gpu_mask_cline[gpu_id],
gpu_fm_inv[gpu_id],
scope='train_loss')
# reuse variables
tf.get_variable_scope().reuse_variables()
# collect gradients and every loss
tower_grads.append(opt.compute_gradients(train_loss))
tower_total_losses.append(train_loss)
tower_data_losses.append(train_data_loss)
tower_smooth_losses.append(train_smooth_loss)
# TensorBoard: collect images from GPU 0
if gpu_id == 0:
gpu0_npr_lines_imgs = gpu_npr_lines[gpu_id]
gpu0_logit_f_imgs = logit_f
gpu0_gt_f_imgs = gpu_gt_field[gpu_id]
gpu0_shape_mask_imgs = gpu_mask_shape[gpu_id]
gpu0_shape_cline_imgs = gpu_mask_cline[gpu_id]
gpu0_ds_imgs = gpu_ds[gpu_id]
gpu0_fm_imgs = gpu_fm[gpu_id]
gpu0_fm_inv_imgs = gpu_fm_inv[gpu_id]
gpu0_selm_imgs = gpu_selm[gpu_id]
gpu0_vdotn_imgs = gpu_vdotn[gpu_id]
tower_loss_collected.append(tower_total_losses)
tower_loss_collected.append(tower_data_losses)
tower_loss_collected.append(tower_smooth_losses)
# Solver
with tf.name_scope('solve') as _:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
grads = average_gradient(tower_grads)
averaged_losses = average_losses(tower_loss_collected)
apply_gradient_op = opt.apply_gradients(grads)
train_op = tf.group(apply_gradient_op)
# TensorBoard: visualization
train_diff_proto = tf.summary.scalar('Training_TotalLoss',
averaged_losses[0])
train_data_loss_proto = tf.summary.scalar('Traning_DataL1Loss',
averaged_losses[1])
train_smooth_loss_proto = tf.summary.scalar('Training_SmoothL1Loss',
averaged_losses[2])
proto_list = collect_vis_img(gpu0_logit_f_imgs, gpu0_npr_lines_imgs,
gpu0_gt_f_imgs, gpu0_shape_mask_imgs,
gpu0_shape_cline_imgs, gpu0_ds_imgs,
gpu0_fm_imgs, gpu0_selm_imgs, gpu0_vdotn_imgs,
gpu0_fm_inv_imgs)
proto_list.append(train_diff_proto)
proto_list.append(train_data_loss_proto)
proto_list.append(train_smooth_loss_proto)
merged_train = tf.summary.merge(proto_list)
return merged_train, train_op, averaged_losses[0], \
[npr_lines, ds, fm, fm_inv, gt_field, cline_mask_inv, mask_shape, mask_2d, selm, vdotn]
def train_procedure(net, train_records, reg_weight):
nb_gpus = hyper_params['nb_gpus']
# Load data
with tf.name_scope('train_inputs') as _:
bSize = hyper_params['batchSize'] * nb_gpus
nbThreads = hyper_params['nbThreads'] * nb_gpus
reader = SketchReader(tfrecord_list=train_records,
raw_size=[256, 256, 25],
shuffle=True,
num_threads=nbThreads,
batch_size=bSize)
raw_input = reader.next_batch()
npr_lines, ds, _, fm, fm_inv, gt_normal, gt_depth, gt_field, mask_cline_inv, mask_shape, mask_ds, _, \
_, mask2d, selm, ndotv = net.cook_raw_inputs(raw_input)
# initialize optimizer
opt = tf.train.AdamOptimizer()
# split data
with tf.name_scope('divide_data'):
gpu_npr_lines = tf.split(npr_lines, nb_gpus, axis=0)
gpu_ds = tf.split(ds, nb_gpus, axis=0)
gpu_fm = tf.split(fm, nb_gpus, axis=0)
gpu_fm_inv = tf.split(fm_inv, nb_gpus, axis=0)
gpu_selm = tf.split(selm, nb_gpus, axis=0)
gpu_ndotv = tf.split(ndotv, nb_gpus, axis=0)
gpu_gt_normal = tf.split(gt_normal, nb_gpus, axis=0)
gpu_gt_depth = tf.split(gt_depth, nb_gpus, axis=0)
gpu_gt_field = tf.split(gt_field, nb_gpus, axis=0)
gpu_mask_shape = tf.split(mask_shape, nb_gpus, axis=0)
gpu_mask_ds = tf.split(mask_ds, nb_gpus, axis=0)
gpu_cline_inv = tf.split(mask_cline_inv, nb_gpus, axis=0)
gpu_mask2d = tf.split(mask2d, nb_gpus, axis=0)
tower_grads = []
tower_loss_collected = []
tower_total_losses = []
tower_d_losses = []
tower_n_losses = []
tower_ds_losses = []
tower_reg_losses = []
tower_abs_d_losses = []
tower_abs_n_losses = []
tower_omega_losses = []
# TensorBoard: images
gpu0_npr_lines_imgs = None
gpu0_gt_ds_imgs = None
gpu0_logit_d_imgs = None
gpu0_logit_n_imgs = None
gpu0_logit_c_imgs = None
gpu0_logit_f_imgs = None
gpu0_gt_normal_imgs = None
gpu0_gt_depth_imgs = None
gpu0_shape_mask_imgs = None
gpu0_gt_field_imgs = None
gpu0_mask_cline_inv_imgs = None
gpu0_fm_imgs = None
gpu0_fm_inv_imgs = None
gpu0_selm_imgs = None
gpu0_ndotv_imgs = None
with tf.variable_scope(tf.get_variable_scope()):
for gpu_id in range(nb_gpus):
with tf.device('gpu:%d' % gpu_id):
with tf.name_scope('tower_%s' % gpu_id) as _:
# Network forward
logit_f, _ = net.load_field_net(gpu_npr_lines[gpu_id],
gpu_mask2d[gpu_id],
gpu_ds[gpu_id],
gpu_fm[gpu_id],
gpu_selm[gpu_id],
gpu_ndotv[gpu_id],
hyper_params['rootFt'],
is_training=False)
logit_d, logit_n, logit_c, _ = net.load_GeomNet(
gpu_npr_lines[gpu_id],
gpu_ds[gpu_id],
gpu_mask2d[gpu_id],
gpu_fm[gpu_id],
gpu_selm[gpu_id],
gpu_ndotv[gpu_id],
logit_f,
gpu_cline_inv[gpu_id],
hyper_params['rootFt'],
is_training=True)
# Training loss
train_loss, train_d_loss, train_n_loss, train_ds_loss, train_reg_loss, train_real_dloss, \
train_real_nloss, train_omega_loss = loss(logit_d,
logit_n,
logit_c,
gpu_gt_normal[gpu_id],
gpu_gt_depth[gpu_id],
gpu_mask_shape[gpu_id],
gpu_mask_ds[gpu_id],
gpu_cline_inv[gpu_id],
reg_weight,
gpu_fm_inv[gpu_id],
scope='train_loss')
# reuse variables
tf.get_variable_scope().reuse_variables()
# collect gradients and every loss
cur_trainable_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='SASMFGeoNet')
tower_grads.append(
opt.compute_gradients(train_loss,
var_list=cur_trainable_vars))
tower_total_losses.append(train_loss)
tower_d_losses.append(train_d_loss)
tower_n_losses.append(train_n_loss)
tower_ds_losses.append(train_ds_loss)
tower_reg_losses.append(train_reg_loss)
tower_abs_d_losses.append(train_real_dloss)
tower_abs_n_losses.append(train_real_nloss)
tower_omega_losses.append(train_omega_loss)
# TensorBoard: collect images from GPU 0
if gpu_id == 0:
gpu0_npr_lines_imgs = gpu_npr_lines[gpu_id]
gpu0_gt_ds_imgs = gpu_ds[gpu_id]
gpu0_logit_d_imgs = logit_d
gpu0_logit_n_imgs = logit_n
gpu0_logit_c_imgs = logit_c
gpu0_logit_f_imgs = logit_f
gpu0_gt_normal_imgs = gpu_gt_normal[gpu_id]
gpu0_gt_depth_imgs = gpu_gt_depth[gpu_id]
gpu0_shape_mask_imgs = gpu_mask_shape[gpu_id]
gpu0_gt_field_imgs = gpu_gt_field[gpu_id]
gpu0_mask_cline_inv_imgs = gpu_cline_inv[gpu_id]
gpu0_fm_imgs = gpu_fm[gpu_id]
gpu0_fm_inv_imgs = gpu_fm_inv[gpu_id]
gpu0_selm_imgs = gpu_selm[gpu_id]
gpu0_ndotv_imgs = gpu_ndotv[gpu_id]
tower_loss_collected.append(tower_total_losses)
tower_loss_collected.append(tower_d_losses)
tower_loss_collected.append(tower_n_losses)
tower_loss_collected.append(tower_ds_losses)
tower_loss_collected.append(tower_reg_losses)
tower_loss_collected.append(tower_abs_d_losses)
tower_loss_collected.append(tower_abs_n_losses)
tower_loss_collected.append(tower_omega_losses)
# Solve
with tf.name_scope('solve') as _:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS,
scope='.*SASMFGeoNet.*')
with tf.control_dependencies(update_ops):
grads = average_gradient(tower_grads)
averaged_losses = average_losses(tower_loss_collected)
apply_gradient_op = opt.apply_gradients(grads)
train_op = tf.group(apply_gradient_op)
# TensorBoard: visualization
train_diff_proto = tf.summary.scalar('Training_TotalLoss',
averaged_losses[0])
train_diff_d_proto = tf.summary.scalar('Training_DepthL2Loss',
averaged_losses[1])
train_diff_n_proto = tf.summary.scalar('Training_NormalL2Loss',
averaged_losses[2])
train_diff_ds_proto = tf.summary.scalar('Training_DepthSampleL2Loss',
averaged_losses[3])
train_diff_reg_proto = tf.summary.scalar('Training_RegL2Loss',
averaged_losses[4])
train_diff_reald_proto = tf.summary.scalar('Training_RealDLoss',
averaged_losses[5])
train_diff_realn_proto = tf.summary.scalar('Training_RealNLoss',
averaged_losses[6])
train_diff_omega_proto = tf.summary.scalar('Training_OmegaLoss',
averaged_losses[7])
proto_list = collect_vis_img(
gpu0_logit_d_imgs, gpu0_logit_n_imgs, gpu0_logit_c_imgs,
gpu0_logit_f_imgs, gpu0_npr_lines_imgs, gpu0_gt_normal_imgs,
gpu0_gt_depth_imgs, gpu0_shape_mask_imgs, gpu0_gt_ds_imgs,
gpu0_gt_field_imgs, gpu0_mask_cline_inv_imgs, gpu0_fm_imgs,
gpu0_fm_inv_imgs, gpu0_selm_imgs, gpu0_ndotv_imgs)
proto_list.append(train_diff_proto)
proto_list.append(train_diff_d_proto)
proto_list.append(train_diff_n_proto)
proto_list.append(train_diff_ds_proto)
proto_list.append(train_diff_reg_proto)
proto_list.append(train_diff_reald_proto)
proto_list.append(train_diff_realn_proto)
proto_list.append(train_diff_omega_proto)
merged_train = tf.summary.merge(proto_list)
return merged_train, train_op, averaged_losses[0]