def evaluate_model(
model,
num_views,
path,
device,
args,
normalise_scale=1,
):
import pickle
was_training = model.training
model.eval()
with torch.no_grad():
m = pickle.load(
open('models/basicModel_%s_lbs_10_207_0_v1.0.0.pkl' %
args.gender[0]))
csv_name = os.path.join(path, 'real_male.csv')
real_height = [1.760837, 1.696549, 1.779879, 1.766942, 1.780513]
# Check real human
with open(csv_name, mode='w') as csv_file:
for i in range(0, args.dataset_size):
inputs = torch.FloatTensor([])
for k in range(0, num_views): # go through the views
img_name = os.path.join(path,
'%d_%d.png' % (i, k)) # image
img = Image.open(img_name)
img = ToTensor()(img)[0]
'''
for x in range(img.shape[0]):
for y in range(img.shape[1]):
if img[x][y] > 0.02:
img[x][y] = 1
'''
img.unsqueeze_(0)
img = torch.cat((img, img, img), 0)
img.unsqueeze_(0)
inputs = torch.cat((inputs, img), 0)
inputs = inputs.to(device)
par_prd = model(inputs)
batch = par_prd.shape[0]
rots, poses, betas = decompose_par(par_prd)
mesh_prd = par_to_mesh(args.gender, rots, poses, betas)
X, Y, Z = [
mesh_prd[:, :, 0], mesh_prd[:, :, 1], mesh_prd[:, :, 2]
]
h_prd, w_prd, c_prd, n_prd, a_prd = vertex2measurements(
X, Y, Z)
vertices = torch.reshape(mesh_prd, (batch, -1))
ratio = real_height[i] / h_prd
csv_writer = csv.writer(csv_file,
delimiter=',',
lineterminator='\n')
csv_writer.writerow([i, 'Height', float(real_height[i])])
csv_writer.writerow([
i,
'Waist',
float(w_prd * ratio),
])
csv_writer.writerow([
i,
'Chest',
float(c_prd * ratio),
])
csv_writer.writerow([i, 'Neck', float(n_prd * ratio)])
csv_writer.writerow([
i,
'Arm',
float(a_prd * ratio),
])
csv_writer.writerow(['\n'])
# rendering
vertices_num = 6890
batch = 1
img_height = 264
img_width = 192
# frontal view
vertices = mesh_prd
name = '%d_0_prd' % i
v = vertices.squeeze().to("cpu").numpy()
angle = np.pi
axis = np.array([0, 0, 1])
v = transpose_mesh(v, angle, axis)
angle = np.pi
axis = np.array([0, 1, 0])
v = transpose_mesh(v, angle, axis)
v_0 = scale_mesh(v, 400, 400, vertices_num=vertices_num)
mesh2Image(v_0,
m['f'],
batch,
path,
name,
400,
400,
vertices_num=vertices_num)
image = Image.open(join(path, name + '.png'))
output = scale_fixed_height(image, img_height, img_width)
output_name = join(path, name + '.png')
output.save(output_name)
# side view
name = '%d_1_prd' % i
angle = np.pi / 2
axis = np.array([0, -1, 0])
v = transpose_mesh(v, angle, axis)
v_1 = scale_mesh(v, 400, 400, vertices_num=vertices_num)
mesh2Image(v_1,
m['f'],
batch,
path,
name,
400,
400,
vertices_num=vertices_num)
image = Image.open(join(path, name + '.png'))
output = scale_fixed_height(image, img_height, img_width)
output_name = join(path, name + '.png')
output.save(output_name)
# mesh generator
outmesh_path = join(path, '%d.obj' % i)
with open(outmesh_path, 'w') as fp:
for v in mesh_prd[0]:
#fp.write( 'v %f %f %f\n' % ( float(v[0]),float(v[1]), float(v[2])) )
fp.write('v %f %f %f\n' % (v[0], v[1], v[2]))
for f in m[
'f'] + 1: # Faces are 1-based, not 0-based in obj files
fp.write('f %d %d %d\n' % (f[0], f[1], f[2]))
csv_file.close()
model.train(mode=was_training)
def main():
args = parse_args()
#args.dataset_size = 10
gender = 'male'
print('-----------------------------------------------------------')
print 'Gender: ', gender
m = pickle.load(
open('models/basicModel_%s_lbs_10_207_0_v1.0.0.pkl' % gender[0]))
dataset_size = args.dataset_size
print 'Dataset range:', args.start_num, ' - ', dataset_size + args.start_num - 1
parent_dic = '/home/yifu/workspace/data/synthetic/noisy/val'
print 'Data Path: ', parent_dic
device = torch.device("cuda:%d" %
args.gpu if torch.cuda.is_available() else "cpu")
torch.cuda.set_device(device)
print('-----------------------------------------------------------')
if raw_input('Confirm the above setting? (yes/no): ') != 'yes':
print('Terminated')
exit()
print 'Data generation starts'
print('------------------------')
for i in range(0, dataset_size):
i = i + args.start_num
rots = (torch.zeros(1, 3)).cuda() # global rotation
poses = (torch.zeros(1, 23, 3)).cuda() # joints
variation = 3
betas = (torch.rand(1, 10) * variation * 2 -
variation).cuda() # shapes
poses[0][17 - 1][2] = np.pi / 4 # right shoulder joint
poses[0][16 - 1][2] = -np.pi / 4 # left shoulder joint
variation = 0.3
poses[0][17 - 1][2] += (torch.rand(1)[0] * variation * 2 -
variation).cuda()
poses[0][16 - 1][2] = -poses[0][17 - 1][2]
variation = 0.15
poses[0][16 - 1][2] += (torch.rand(1)[0] * variation * 2 -
variation).cuda()
poses[0][1 - 1][2] = np.pi / 40
variation = np.pi / 40
poses[0][1 - 1][2] += (torch.rand(1)[0] * variation * 2 -
variation).cuda()
poses[0][2 - 1][2] = -np.pi / 40
variation = np.pi / 40
poses[0][2 - 1][2] += (torch.rand(1)[0] * variation * 2 -
variation).cuda()
variation = 0.03
for j in range(0, 23):
for k in range(0, 3):
poses[0][j][k] += torch.rand(1)[0] * variation * 2 - variation
vertices = par_to_mesh(gender, rots, poses, betas)
# rendering
vertices_num = 6890
path = parent_dic
batch = 1
img_height = 264
img_width = 192
# frontal view
name = '%d_0' % i
v = vertices.squeeze().to("cpu").numpy()
angle = np.pi
axis = np.array([0, 0, 1])
v = transpose_mesh(v, angle, axis)
angle = np.pi
axis = np.array([0, 1, 0])
v = transpose_mesh(v, angle, axis)
v_0 = scale_mesh(v, 400, 400, vertices_num=vertices_num)
mesh2Image(v_0,
m['f'],
batch,
path,
name,
400,
400,
vertices_num=vertices_num)
image = Image.open(join(path, name + '.png'))
output = scale_fixed_height(image, img_height, img_width)
output_name = join(path, name + '.png')
output.save(output_name)
# side view
name = '%d_1' % i
angle = np.pi / 2
axis = np.array([0, -1, 0])
v = transpose_mesh(v, angle, axis)
v_1 = scale_mesh(v, 400, 400, vertices_num=vertices_num)
mesh2Image(v_1,
m['f'],
batch,
path,
name,
400,
400,
vertices_num=vertices_num)
image = Image.open(join(path, name + '.png'))
output = scale_fixed_height(image, img_height, img_width)
output_name = join(path, name + '.png')
output.save(output_name)
# pickle
poses = torch.reshape(poses, (1, -1))
parameters = torch.cat((rots[0], poses[0], betas[0]), 0).to("cpu")
pickle.dump(parameters, open('%s/%d' % (path, i), 'wb'))
'''
csv_writer.writerow([i, 'Arm', arm])
csv_writer.writerow(['\n'])
batch = 1
height = 400
width = 400
mesh = vertices
name = name[:-4]
#face = m.f
faces = faces.astype(np.uint32)
name = name + '_0'
angle = np.pi
axis = np.array([0, 0, 1])
mesh = transpose_mesh(mesh, angle, axis)
angle = np.pi
axis = np.array([0, 1, 0])
mesh = transpose_mesh(mesh, angle, axis)
mesh_0 = scale_mesh(mesh, height, width, vertices_num = vertices_n)
mesh2Image(mesh_0, faces, batch, path, name, height, width, vertices_num = vertices_n)
name = name[:-2] + '_1'
angle = -np.pi/2
axis = np.array([0, 1, 0])
mesh = transpose_mesh(mesh, angle, axis)
mesh_1 = scale_mesh(mesh, height, width,vertices_num = vertices_n)
mesh2Image(mesh_1, faces, batch, path, name, height, width,vertices_num = vertices_n)
h, w, c, n, a = vertex2measurements(X, Y, Z)
#vertices = par_to_mesh(gender, rots, poses, betas)
# rendering
parent_dic = '/home/yifu/workspace/data_smpl/test'
vertices_num = 6890
path = parent_dic
batch = 1
# frontal view
name = 'test_0'
v = vertices.squeeze().to("cpu").numpy()
angle = np.pi
axis = np.array([0, 0, 1])
v = transpose_mesh(v, angle, axis)
angle = np.pi
axis = np.array([0, 1, 0])
v = transpose_mesh(v, angle, axis)
v_0 = scale_mesh(v, 400, 400, vertices_num=vertices_num)
mesh2Image(v_0, m['f'], batch, path, name, 400, 400, vertices_num=vertices_num)
# side view
name = 'test_1'
angle = np.pi / 2
axis = np.array([0, -1, 0])
v = transpose_mesh(v, angle, axis)
v_1 = scale_mesh(v, 400, 400, vertices_num=vertices_num)
mesh2Image(v_1, m['f'], batch, path, name, 400, 400, vertices_num=vertices_num)
# mesh generator
outmesh_path = join(path, '%d.obj' % i)