def deal(x):
v1 = x[6:9]
x[6:9] = normalize(v1)
v1 = x[6:9]
v2 = x[9:12]
v2 = v2 - proj(v2, v1)
x[9:12] = normalize(v2)
#
v1 = x[18:21]
x[18:21] = normalize(v1)
v1 = x[18:21]
v2 = x[21:24]
v2 = v2 - proj(v2, v1)
x[21:24] = normalize(v2)
def animate(i, config, net, optim, data):
print(i, '/', iternum)
net.train()
out = net(data)
loss = config.loss(data, out)
optim.zero_grad()
loss['overall'].backward()
optim.step()
print(loss['overall'].data.cpu().numpy())
net.eval()
with torch.no_grad():
out = net(data)
box3d_src = data[1].data.cpu().numpy()
box3d_tgt = data[2].data.cpu().numpy()
center = data[3].data.cpu().numpy()
r = data[4].data.cpu().numpy()
gts = data[5].data.cpu().numpy()
y = out['y'].data.cpu().numpy()
map = out['dmap'].permute(0, 2, 3, 1).data.cpu().numpy()
#
num = box3d_src.shape[0]
col = int(np.sqrt(num))
row = num // col
for ri in range(row):
for cj in range(col):
ni = ri * col + cj
ymap = y[ni, ...]
ymap *= np.pi
ymap[1] *= 2
c3d = recon(center[ni, ...], r[ni, ...], ymap)
c2d = proj(mv(c3d))
pv[4 * ni + 0].set_data(c2d[:, 0], c2d[:, 1])
im1 = norm_im(map[ni, :, :, :3])
pv[4 * ni + 1].set_array(im1)
im2 = norm_im(map[ni, :, :, 3:6])
pv[4 * ni + 2].set_array(im2)
pv[4 * ni + 3].set_data(c3d[:, 0], c3d[:, 1])
pv[4 * ni + 3].set_3d_properties(c3d[:, 2])
if i == iternum - 1:
exit()
return pv
def run(**kwargs):
print('a')
#get configuration
try:
config = importlib.import_module('config.'+kwargs['config']);
opt = config.__dict__;
for k in kwargs.keys():
if not kwargs[k] is None:
opt[k] = kwargs[k];
except Exception as e:
print(e);
traceback.print_exc();
exit();
#get network
try:
m = importlib.import_module('net.'+opt['net']);
net = m.Net(**opt);
if torch.cuda.is_available():
net = net.cuda();
except Exception as e:
print(e);
traceback.print_exc();
exit();
#get dataset
try:
m = importlib.import_module('util.dataset.'+opt['dataset']);
train_data = m.Data(opt,True);
val_data = m.Data(opt,False);
train_load = DataLoader(train_data,batch_size=opt['batch_size'],shuffle=True,num_workers=opt['workers']);
val_load = DataLoader(val_data,batch_size=opt['batch_size'],shuffle=False,num_workers=opt['workers']);
except Exception as e:
print(e);
traceback.print_exc();
exit();
if opt['model']!='':
partial_restore(net,opt['model']);
print("Previous weights loaded");
if 'train' in opt['user_key']:
load = train_load;
else:
load = val_load;
if opt['model']!='':
outdir = os.path.dirname(opt['model'])+os.sep+'view_'+opt['user_key'];
if not os.path.exists(outdir):
os.mkdir(outdir);
for i, data in enumerate(load,0):
print(i,'/',len(train_data)//opt['batch_size']);
data2cuda(data);
net.eval();
with torch.no_grad():
out = net(data);
img = data[0].data.cpu().numpy();
box2d_src = data[1].data.cpu().numpy();
box3d_src = data[2].data.cpu().numpy();
box2d_tgt = data[3].data.cpu().numpy();
box3d_tgt = data[4].data.cpu().numpy();
r = data[5].data.cpu().numpy();
gts = data[6].data.cpu().numpy();
y = out['y'].data.cpu().numpy();
tri = box_face;
num = box3d_src.shape[0];
col = 4;
row = num // col;
for ri in range(row):
for cj in range(col):
ni = ri*col+cj;
fig = plt.figure(figsize=(48,16));
#
ax = fig.add_subplot(132,projection='3d');
ax.view_init(elev=20, azim=0)
ax.plot_trisurf(box3d_tgt[ni,...,0],box3d_tgt[ni,...,1],tri,box3d_tgt[ni,...,2],color=(0,0,1,0.1));
ax.plot_trisurf(box3d_src[ni,...,0],box3d_src[ni,...,1],tri,box3d_src[ni,...,2],color=(0,1,0,0.1));
ygt = gts[ni,...];
ygt *= np.pi;
ygt[1] *= 2;
c3d1 = recon(box3d_src[ni,...],r[ni,...],ygt);
ax.plot(c3d1[:,0],c3d1[:,1],c3d1[:,2],color='k');
ymap = y[ni,...];
ymap *= np.pi;
ymap[1] *= 2;
c3d2 = recon(box3d_src[ni,...],r[ni,...],ymap);
ax.plot(c3d2[:,0],c3d2[:,1],c3d2[:,2],color='r');
ax.scatter(box3d_src[ni,0:4,0],box3d_src[ni,0:4,1],box3d_src[ni,0:4,2],color='b',marker='*');
ax.scatter(box3d_src[ni,4:8,0],box3d_src[ni,4:8,1],box3d_src[ni,4:8,2],color='c',marker='*');
ax.scatter(box3d_tgt[ni,0:4,0],box3d_tgt[ni,0:4,1],box3d_tgt[ni,0:4,2],color='k',marker='x');
ax.scatter(box3d_tgt[ni,4:8,0],box3d_tgt[ni,4:8,1],box3d_tgt[ni,4:8,2],color='r',marker='x');
ax.set_aspect('equal', adjustable='box');
#
ax = fig.add_subplot(133,projection='3d');
ax.view_init(elev=20, azim=90)
ax.plot_trisurf(box3d_tgt[ni,...,0],box3d_tgt[ni,...,1],tri,box3d_tgt[ni,...,2],color=(0,0,1,0.1));
ax.plot_trisurf(box3d_src[ni,...,0],box3d_src[ni,...,1],tri,box3d_src[ni,...,2],color=(0,1,0,0.1));
ax.plot(c3d1[:,0],c3d1[:,1],c3d1[:,2],color='k');
ax.plot(c3d2[:,0],c3d2[:,1],c3d2[:,2],color='r');
ax.scatter(box3d_src[ni,0:4,0],box3d_src[ni,0:4,1],box3d_src[ni,0:4,2],color='b',marker='*');
ax.scatter(box3d_src[ni,4:8,0],box3d_src[ni,4:8,1],box3d_src[ni,4:8,2],color='c',marker='*');
ax.scatter(box3d_tgt[ni,0:4,0],box3d_tgt[ni,0:4,1],box3d_tgt[ni,0:4,2],color='k',marker='x');
ax.scatter(box3d_tgt[ni,4:8,0],box3d_tgt[ni,4:8,1],box3d_tgt[ni,4:8,2],color='r',marker='x');
ax.set_aspect('equal', adjustable='box');
#
ax = fig.add_subplot(131);
ax.set_aspect('equal', adjustable='box');
ax.imshow(img[ni,...]);
ax.scatter(box2d_src[ni,0:4,0],box2d_src[ni,0:4,1],color='b',marker='*');
ax.scatter(box2d_src[ni,4:8,0],box2d_src[ni,4:8,1],color='c',marker='*');
ax.scatter(box2d_tgt[ni,0:4,0],box2d_tgt[ni,0:4,1],color='k',marker='x');
ax.scatter(box2d_tgt[ni,4:8,0],box2d_tgt[ni,4:8,1],color='r',marker='x');
ax.set_aspect('equal', adjustable='box');
c2d1 = proj(mv(c3d1));
c2d2 = proj(mv(c3d2));
ax.plot(c2d1[:,0],c2d1[:,1],color='k');
ax.plot(c2d2[:,0],c2d2[:,1],color='r');
if opt['model']!='':
plt.savefig(os.path.join(outdir,"_%04d_%04d.png"%(i,ni)));
if opt['ply']:
plt.show();
plt.close(fig);
#run the code
'''
def run(**kwargs):
global iternum
#get configuration
try:
config = importlib.import_module('config.' + kwargs['config'])
opt = config.__dict__
for k in kwargs.keys():
if not kwargs[k] is None:
opt[k] = kwargs[k]
except Exception as e:
print(e)
traceback.print_exc()
exit()
iternum = opt['nepoch']
#get network
try:
m = importlib.import_module('net.' + opt['net'])
net = m.Net(**opt)
if torch.cuda.is_available():
net = net.cuda()
except Exception as e:
print(e)
traceback.print_exc()
exit()
#get dataset
try:
m = importlib.import_module('util.dataset.' + opt['dataset'])
train_data = m.Data(opt, True)
val_data = m.Data(opt, False)
train_load = DataLoader(train_data,
batch_size=opt['batch_size'],
shuffle=True,
num_workers=opt['workers'])
val_load = DataLoader(val_data,
batch_size=opt['batch_size'],
shuffle=False,
num_workers=opt['workers'])
except Exception as e:
print(e)
traceback.print_exc()
exit()
if opt['model'] != '':
partial_restore(net, opt['model'])
print("Previous weights loaded")
if 'train' in opt['user_key']:
load = train_load
print(train_data.datapath[0])
else:
load = val_load
print(val_data.datapath[0])
if opt['model'] != '':
outdir = os.path.dirname(
opt['model']) + os.sep + 'view_' + opt['user_key']
if not os.path.exists(outdir):
os.mkdir(outdir)
for i, data in enumerate(load, 0):
data2cuda(data)
d = data
#
net.eval()
with torch.no_grad():
out = net(data)
#
img = data[0].data.cpu().numpy()
box_src = data[1].data.cpu().numpy()
box_tgt = data[2].data.cpu().numpy()
center = data[3].data.cpu().numpy()
r = data[4].data.cpu().numpy()
gt = data[5].data.cpu().numpy()
yout = out['y'].data.cpu().numpy()
map = out['dmap'].permute(0, 2, 3, 1).data.cpu().numpy()
#run the code
#optim = eval('optim.'+opt['optim'])(config.parameters(net),lr=opt['lr'],weight_decay=opt['weight_decay']);
tri = box_face
global pv
num = box_src.shape[0]
col = int(np.sqrt(num))
row = num // col
for ri in range(row):
for cj in range(col):
ni = ri * col + cj
fig = plt.figure(figsize=(48, 16))
#===========================================
y = gt[ni, ...].copy()
y *= np.pi
y[1] *= 2
c3d = recon(center[ni, ...], r[ni, ...], y)
#===========================================
ymap = yout[ni, ...]
ymap *= np.pi
ymap[1] *= 2
c3do = recon(center[ni, ...], r[ni, ...], ymap)
c2do = proj(mv(c3do))
#===========================================
ax = fig.add_subplot(1, 5, 1)
ax.imshow(img[ni, ...])
c2d = proj(mv(c3d))
line = ax.plot(c2do[:, 0], c2do[:, 1], color='r')
pv.extend(line)
ax.plot(c2d[:, 0], c2d[:, 1], color='k')
#===========================================
ax = fig.add_subplot(1, 5, 2)
im1 = norm_im(map[ni, :, :, :3])
im1 = ax.imshow(im1)
pv.append(im1)
#===========================================
ax = fig.add_subplot(1, 5, 3)
im2 = norm_im(map[ni, :, :, 3:])
im2 = ax.imshow(im2)
pv.append(im2)
#===========================================
ax = fig.add_subplot(1, 5, 4, projection='3d')
ax.view_init(elev=20, azim=90)
ax.set_aspect('equal', adjustable='box')
ax.plot_trisurf(box_tgt[ni, ..., 0],
box_tgt[ni, ..., 1],
tri,
box_tgt[ni, ..., 2],
color=(0, 0, 1, 0.1))
ax.plot_trisurf(box_src[ni, ..., 0],
box_src[ni, ..., 1],
tri,
box_src[ni, ..., 2],
color=(0, 1, 0, 0.1))
line = ax.plot(c3do[:, 0], c3do[:, 1], c3do[:, 2], color='r')
#==========================================
pv.extend(line)
ax.plot(c3d[:, 0], c3d[:, 1], c3d[:, 2], color='k')
#==========================================
ax = fig.add_subplot(1, 5, 5, projection='3d')
ax.set_aspect('equal', adjustable='box')
ax.plot_trisurf(box_tgt[ni, ..., 0],
box_tgt[ni, ..., 1],
tri,
box_tgt[ni, ..., 2],
color=(0, 0, 1, 0.1))
ax.plot_trisurf(box_src[ni, ..., 0],
box_src[ni, ..., 1],
tri,
box_src[ni, ..., 2],
color=(0, 1, 0, 0.1))
line = ax.plot(c3do[:, 0], c3do[:, 1], c3do[:, 2], color='r')
#===========================================
pv.extend(line)
ax.plot(c3d[:, 0], c3d[:, 1], c3d[:, 2], color='k')
#===========================================
if opt['model'] != '':
plt.savefig(
os.path.join(outdir, "_%04d_%04d.png" % (i, ni)))
if opt['ply']:
plt.show()
plt.close(fig)
#============
#run the code
'''
def run(**kwargs):
global iternum
#get configuration
try:
config = importlib.import_module('config.' + kwargs['config'])
opt = config.__dict__
for k in kwargs.keys():
if not kwargs[k] is None:
opt[k] = kwargs[k]
except Exception as e:
print(e)
traceback.print_exc()
exit()
iternum = opt['nepoch']
#get network
try:
m = importlib.import_module('net.' + opt['net'])
net = m.Net(**opt)
if torch.cuda.is_available():
net = net.cuda()
except Exception as e:
print(e)
traceback.print_exc()
exit()
#get dataset
try:
m = importlib.import_module('util.dataset.' + opt['dataset'])
train_data = m.Data(opt, True)
val_data = m.Data(opt, False)
train_load = DataLoader(train_data,
batch_size=opt['batch_size'],
shuffle=True,
num_workers=opt['workers'])
val_load = DataLoader(val_data,
batch_size=opt['batch_size'],
shuffle=False,
num_workers=opt['workers'])
except Exception as e:
print(e)
traceback.print_exc()
exit()
for i, data in enumerate(train_load, 0):
data2cuda(data)
d = data
break
#
img = data[0].data.cpu().numpy()
box_src = data[1].data.cpu().numpy()
box_tgt = data[2].data.cpu().numpy()
center = data[3].data.cpu().numpy()
r = data[4].data.cpu().numpy()
gt = data[5].data.cpu().numpy()
#run the code
optim = eval('optim.' + opt['optim'])(config.parameters(net),
lr=opt['lr'],
weight_decay=opt['weight_decay'])
tri = box_face
global pv
num = box_src.shape[0]
col = int(np.sqrt(num))
row = num // col
for ri in range(row):
for cj in range(col):
ni = ri * col + cj
y = gt[ni, ...].copy()
y *= np.pi
y[1] *= 2
c3d = recon(center[ni, ...], r[ni, ...], y)
#===========================================
ax = fig.add_subplot(row, col * 4, 4 * ni + 1)
ax.imshow(img[ni, ...])
c2d = proj(mv(c3d))
line = ax.plot(c2d[:, 0], c2d[:, 1], color='r')
pv.extend(line)
ax.plot(c2d[:, 0], c2d[:, 1], color='k')
#===========================================
ax = fig.add_subplot(row, col * 4, 4 * ni + 2)
im1 = ax.imshow(img[ni, ...])
pv.append(im1)
#===========================================
ax = fig.add_subplot(row, col * 4, 4 * ni + 3)
im2 = ax.imshow(img[ni, ...])
pv.append(im2)
ax = fig.add_subplot(row, col * 4, 4 * ni + 4, projection='3d')
ax.axis('equal')
ax.plot_trisurf(box_tgt[ni, ..., 0],
box_tgt[ni, ..., 1],
tri,
box_tgt[ni, ..., 2],
color=(0, 0, 1, 0.1))
ax.plot_trisurf(box_src[ni, ..., 0],
box_src[ni, ..., 1],
tri,
box_src[ni, ..., 2],
color=(0, 1, 0, 0.1))
line = ax.plot(c3d[:, 0], c3d[:, 1], c3d[:, 2], color='r')
pv.extend(line)
ax.plot(c3d[:, 0], c3d[:, 1], c3d[:, 2], color='k')
#run the code
animfunc = partial(animate, config=config, net=net, optim=optim, data=data)
anim = animation.FuncAnimation(fig,
animfunc,
init_func=init,
frames=iternum,
interval=30,
blit=False)
plt.show()
'''