def call(self, inp):
images, vertexlabel, Js_in = inp
out_dict = {}
images = [
tf.Variable(x, dtype=tf.float32, trainable=False) for x in images
]
vertexlabel = tf.cast(tf.Variable(vertexlabel, trainable=False),
tf.int32)
if FACE:
Js = [Lambda(lambda j: j[:, :25])(J) for J in Js_in]
else:
Js = [
self.flatten(
tf.cast(tf.Variable(x, trainable=False), tf.float32))
for x in Js_in
]
with tf.device('/gpu:1'):
lat_codes = [self.top_([q, j]) for q, j in zip(images, Js)]
latent_code_offset = self.avg([q[0] for q in lat_codes])
latent_code_betas = self.avg([q[1] for q in lat_codes])
latent_code_pose = [
tf.concat([q[1], x], axis=-1) for q, x in zip(lat_codes, Js)
]
with tf.device('/gpu:2'):
latent_code_betas = self.lat_betas(latent_code_betas)
betas = self.betas(latent_code_betas)
latent_code_pose = [self.lat_pose(x) for x in latent_code_pose]
pose_trans_init = tf.tile(tf.expand_dims(self.pose_trans, 0),
(K.int_shape(betas)[0], 1))
poses_ = [
self.lat_pose_layer(x) + pose_trans_init
for x in latent_code_pose
]
trans_ = [self.cut_trans(x) for x in poses_]
trans = [la(i) for la, i in zip(self.trans_layers, trans_)]
poses_ = [self.cut_poses(x) for x in poses_]
poses_ = [self.reshape_pose(x) for x in poses_]
poses = [la(i) for la, i in zip(self.pose_layers, poses_)]
##
out_dict['betas'] = betas
for i in range(NUM):
out_dict['pose_{}'.format(i)] = poses[i]
out_dict['trans_{}'.format(i)] = trans[i]
latent_code_offset_ShapeMerged = self.latent_code_offset_ShapeMerged(
latent_code_offset)
latent_code_offset_ShapeMerged = self.latent_code_offset_ShapeMerged_2(
latent_code_offset_ShapeMerged)
garm_model_outputs = [
fe(latent_code_offset_ShapeMerged) for fe in self.garmentModels
]
garment_verts_all = [fe[0] for fe in garm_model_outputs]
garment_pca = [fe[1] for fe in garm_model_outputs]
garment_pca = tf.stack(garment_pca, axis=1)
##
out_dict['pca_verts'] = garment_pca
lis = []
for go, vs in zip(garment_verts_all, self.scatters):
lis.append(vs(go))
garment_verts_all_scattered = tf.stack(lis, axis=-1)
# Get naked smpl to compute garment offsets
zerooooooo = K.zeros_like(garment_verts_all_scattered[..., 0])
pooooooooo = [K.zeros_like(p) for p in poses]
tooooooooo = [K.zeros_like(p) for p in trans]
smpls_base = []
for i, (p, t) in enumerate(zip(pooooooooo, tooooooooo)):
v, _, n, _ = self.smpl(p, betas, t, zerooooooo)
smpls_base.append(v)
if i == 0:
vertices_naked_ = n
# Append Skin offsets
garment_verts_all_scattered = tf.concat([
K.expand_dims(vertices_naked_, -1),
tf.cast(garment_verts_all_scattered, vertices_naked_.dtype)
],
axis=-1)
garment_verts_all_scattered = tf.transpose(
garment_verts_all_scattered, perm=[0, 1, 3, 2])
clothed_verts = tf.batch_gather(garment_verts_all_scattered,
vertexlabel)
clothed_verts = tf.squeeze(
tf.transpose(clothed_verts, perm=[0, 1, 3, 2]))
offsets_ = clothed_verts - vertices_naked_
smpls = []
for i, (p, t) in enumerate(zip(poses, trans)):
v, t, n, _ = self.smpl(p, betas, t, offsets_)
smpls.append(v)
if i == 0:
vertices_naked = n
vertices_tposed = t
Js = [
jl(self.smpl_J([p, betas, t]))
for jl, p, t in zip(self.J_layers, poses, trans)
]
vertices = tf.concat([
tf.expand_dims(smpl, axis=-1) for i, smpl in enumerate(smpls)
],
axis=-1)
##
out_dict['vertices'] = vertices
out_dict['vertices_tposed'] = vertices_tposed
out_dict['vertices_naked'] = vertices_naked
##
out_dict['vertices'] = vertices
out_dict['vertices_tposed'] = vertices_tposed
out_dict['vertices_naked'] = vertices_naked
out_dict['offsets_h'] = offsets_
for i in range(NUM):
out_dict['J_{}'.format(i)] = Js[i]
vert_cols = tf.reshape(
tf.gather(self.colormap, tf.reshape(vertexlabel, (-1, ))),
(-1, config.NVERTS, 3))
renderered_garms_all = []
for view in range(NUM):
renderered_garms_all.append(
render_colored_batch(
vertices[..., view],
self.faces,
vert_cols, # [bat],
IMG_SIZE,
IMG_SIZE,
FOCAL_LENGTH,
CAMERA_CENTER,
np.zeros(3, dtype=np.float32),
num_channels=3))
renderered_garms_all = tf.transpose(renderered_garms_all,
[1, 2, 3, 4, 0])
out_dict['rendered'] = renderered_garms_all
lap = compute_laplacian_diff(vertices_tposed, vertices_naked,
self.faces)
##
out_dict['laplacian'] = lap
return out_dict
def call(self, inp):
images, vertexlabel, Js_in = inp
out_dict = {}
"""
print('\n'*2 + '='*99)
print(len(images)) # 8
print(images[0].shape)# (2, 480, 270, 3)
print('='*99 + '\n'*2 )
#raise Exception("nxb stopped execution here.")
"""
# NoTE: this is the last place I was trying to understand why img1[0] and img1[1] aren't images of the same person from the same view. -nxb
# NOTE: BLB told us that test_network.py runs on 2 people simultaenously
images = [
tf.Variable(x, dtype=tf.float32, trainable=False) for x in images
] # list of 2 images (as tf.Variable()s)
vertexlabel = tf.cast(tf.Variable(vertexlabel, trainable=False),
tf.int32)
if FACE:
Js = [Lambda(lambda j: j[:, :25])(J) for J in Js_in]
else:
Js = [
self.flatten(
tf.cast(tf.Variable(x, trainable=False), tf.float32))
for x in Js_in
]
with tf.device('/gpu:0'):
lat_codes = [self.top_([q, j]) for q, j in zip(images, Js)]
latent_code_offset = self.avg([q[0] for q in lat_codes])
latent_code_betas = self.avg([q[1] for q in lat_codes])
latent_code_pose = [
tf.concat([q[1], x], axis=-1) for q, x in zip(lat_codes, Js)
]
latent_code_betas = self.lat_betas(latent_code_betas)
betas = self.betas(latent_code_betas)
latent_code_pose = [self.lat_pose(x) for x in latent_code_pose]
pose_trans_init = tf.tile(tf.expand_dims(self.pose_trans, 0),
(K.int_shape(betas)[0], 1))
poses_ = [
self.lat_pose_layer(x) + pose_trans_init
for x in latent_code_pose
]
trans_ = [self.cut_trans(x) for x in poses_]
trans = [la(i) for la, i in zip(self.trans_layers, trans_)]
poses_ = [self.cut_poses(x) for x in poses_]
poses_ = [self.reshape_pose(x) for x in poses_]
poses = [la(i) for la, i in zip(self.pose_layers, poses_)]
##
out_dict['betas'] = betas
for i in range(NUM):
out_dict['pose_{}'.format(i)] = poses[i]
out_dict['trans_{}'.format(i)] = trans[i]
latent_code_offset_ShapeMerged = self.latent_code_offset_ShapeMerged(
latent_code_offset)
latent_code_offset_ShapeMerged = self.latent_code_offset_ShapeMerged_2(
latent_code_offset_ShapeMerged)
garm_model_outputs = [
fe(latent_code_offset_ShapeMerged) for fe in self.garmentModels
]
garment_verts_all = [fe[0] for fe in garm_model_outputs]
garment_pca = [fe[1] for fe in garm_model_outputs]
garment_pca = tf.stack(garment_pca, axis=1)
##
out_dict['pca_verts'] = garment_pca
lis = []
for go, vs in zip(garment_verts_all, self.scatters):
lis.append(vs(go))
garment_verts_all_scattered = tf.stack(lis, axis=-1)
## Get naked smpl to compute garment offsets
zerooooooo = K.zeros_like(garment_verts_all_scattered[..., 0])
pooooooooo = [K.zeros_like(p) for p in poses]
tooooooooo = [K.zeros_like(p) for p in trans]
smpls_base = []
for i, (p, t) in enumerate(zip(pooooooooo, tooooooooo)):
v, _, n, _ = self.smpl(p, betas, t, zerooooooo)
smpls_base.append(v)
if i == 0:
vertices_naked_ = n
## Append Skin offsets
garment_verts_all_scattered = tf.concat([
K.expand_dims(vertices_naked_, -1),
tf.cast(garment_verts_all_scattered, vertices_naked_.dtype)
],
axis=-1)
garment_verts_all_scattered = tf.transpose(
garment_verts_all_scattered, perm=[0, 1, 3, 2])
clothed_verts = tf.batch_gather(garment_verts_all_scattered,
vertexlabel)
clothed_verts = tf.squeeze(
tf.transpose(clothed_verts, perm=[0, 1, 3, 2]))
offsets_ = clothed_verts - vertices_naked_
smpls = []
for i, (p, t) in enumerate(zip(poses, trans)):
v, t, n, _ = self.smpl(p, betas, t, offsets_)
smpls.append(v)
if i == 0:
vertices_naked = n
vertices_tposed = t
Js = [
jl(self.smpl_J(p, betas, t))
for jl, p, t in zip(self.J_layers, poses, trans)
]
#Js = [ jl(self.smpl_J([p, betas, t])) for jl, p, t in zip(self.J_layers, poses, trans) ] # original line (pre- Mon Feb 10 20:19:03 EST 2020)
vertices = tf.concat(
[tf.expand_dims(smpl, axis=-1) for i, smpl in enumerate(smpls)],
axis=-1)
##
out_dict['vertices'] = vertices
out_dict['vertices_tposed'] = vertices_tposed
out_dict['vertices_naked'] = vertices_naked
##
out_dict['vertices'] = vertices
out_dict['vertices_tposed'] = vertices_tposed
out_dict['vertices_naked'] = vertices_naked
out_dict['offsets_h'] = offsets_
for i in range(NUM):
out_dict['J_{}'.format(i)] = Js[i]
vert_cols = tf.reshape(
tf.gather(self.colormap, tf.reshape(vertexlabel, (-1, ))),
(-1, config.NVERTS, 3))
renderered_garms_all = []
for view in range(NUM):
renderered_garms_all.append(
render_colored_batch(
vertices[..., view],
self.faces,
vert_cols, # [bat],
IMG_SIZE,
IMG_SIZE,
FOCAL_LENGTH,
CAMERA_CENTER,
np.zeros(3, dtype=np.float32),
num_channels=3))
renderered_garms_all = tf.transpose(renderered_garms_all,
[1, 2, 3, 4, 0])
out_dict['rendered'] = renderered_garms_all
lap = compute_laplacian_diff(vertices_tposed, vertices_naked,
self.faces)
##
out_dict['laplacian'] = lap
return out_dict