def do_train(cur_step, optimizer, sim, model):
model.train()
while True:
steps = 4 * 25 * spf
reset_sim(sim)
st = time.time()
loss, dis, vext = run_sim(steps, sim, model)
en0 = time.time()
# loss = get_loss(steps,sim)
optimizer.zero_grad()
print('step={}'.format(cur_step))
print('forward time={}'.format(en0 - st))
print('loss={}'.format(loss.data))
f.write('step {}: loss={}\n'.format(cur_step, loss.data))
print('step {}: loss={}\n'.format(cur_step, loss.data))
loss.backward()
nn.utils.clip_grad_value_(model.parameters(), 10)
if dis[:2].norm() < 0.5 and dis[2] > 0 and dis[2] < 0.5:
break
en1 = time.time()
print('backward time={}'.format(en1 - en0))
optimizer.step()
#f.write('{}\n'.format(vext.grad));
f.write('{}\n'.format(vext * scalev))
arcsim.delete_mesh(sim.cloths[0].mesh)
# break
torch.save(model.state_dict(), sys.argv[1] + "/3.pt")
def do_train(cur_step,optimizer,sim,vext):
while True:
steps=7*25*spf
reset_sim(sim)
st = time.time()
loss, dis = run_sim(steps, sim, vext)
# loss = loss + regularization(vext)
en0 = time.time()
optimizer.zero_grad()
print('step={}'.format(cur_step))
print('forward time={}'.format(en0-st))
print('loss={}'.format(loss.data))
f.write('step {}: loss={}\n'.format(cur_step, loss.data))
print('step {}: loss={}\n'.format(cur_step, loss.data))
loss.backward()
vext.grad.data.clamp_(-10,10)
if dis[:2].norm() < 0.5 and dis[2]>0 and dis[2]<0.5:
break
en1 = time.time()
print('backward time={}'.format(en1-en0))
print(vext.grad)
optimizer.step()
print(vext*scalev)
f.write('{}\n'.format(vext*scalev));
arcsim.delete_mesh(sim.cloths[0].mesh)
def do_train(cur_step, optimizer, sim, vext):
while True:
steps = 4 * 25 * spf
reset_sim(sim)
st = time.time()
loss, dis = run_sim(steps, sim, vext)
en0 = time.time()
# loss = get_loss(steps,sim)
optimizer.zero_grad()
print('step={}'.format(cur_step))
print('forward time={}'.format(en0 - st))
print('loss={}'.format(loss.data))
f.write('step {}: loss={}\n'.format(cur_step, loss.data))
print('step {}: loss={}\n'.format(cur_step, loss.data))
loss.backward()
vext.grad.data.clamp_(-10, 10)
if dis[:2].norm() < 0.5 and dis[2] > 0 and dis[2] < 0.5: #loss<1e-3:#
break
# dgrad, stgrad, begrad = torch.autograd.grad(loss, [density, stretch, bend])
en1 = time.time()
print('backward time={}'.format(en1 - en0))
print(vext.grad)
optimizer.step()
print(vext * scalev)
#f.write('{}\n'.format(vext.grad));
f.write('{}\n'.format(vext * scalev))
arcsim.delete_mesh(sim.cloths[0].mesh)
def naive_guess(density, stretchori, bend):
return density, stretchori, bend
M = []
for i in range(5, 25):
fstd = sys.argv[2] + '/%04d_00.obj' % (i)
mesh = arcsim.Mesh()
arcsim.load_obj(mesh, fstd)
fstd = sys.argv[2] + '/%04d_00.obj' % (i + 1)
arcsim.load_obj(sim.cloths[0].mesh, fstd)
arcsim.compute_ms_data(sim.cloths[0].mesh)
for n0 in sim.cloths[0].mesh.nodes:
n0.m = torch.zeros([], dtype=torch.float64)
n = len(sim.cloths[0].mesh.nodes)
fext = torch.zeros([n, 3], dtype=torch.float64)
Jext = torch.zeros([n, 3], dtype=torch.float64)
arcsim.add_external_forces(sim.cloths[0], sim.gravity, sim.wind, fext,
Jext)
m = []
i = 0
for n0, n1 in zip(mesh.nodes, sim.cloths[0].mesh.nodes):
a = (n1.v - n0.v) / config['frame_time'] * config['frame_steps']
a[2] += 9.8
if i != 2 or i != 3:
m.append((fext[i].sum() / a.sum()))
i += 1
arcsim.delete_mesh(mesh)
M.append(torch.relu(torch.stack(m).sum()))
print('density={}'.format(torch.stack(M).mean()))
density = torch.stack(M).mean() / scaleden
mesh = arcsim.Mesh()
fstd = sys.argv[2] + '/%04d_00.obj' % (4)
arcsim.load_obj(mesh, fstd)
mini = 1e10
maxi = -1e10
for n0 in mesh.nodes:
mini = min(mini, n0.x[2])
maxi = max(maxi, n0.x[2])
defo = maxi - mini - 1
stre = density * 9.8 / defo * 2
arcsim.delete_mesh(mesh)
print(stre)
stretch = torch.tensor([[1, 0, 1, 0]], dtype=torch.float64).repeat(
[6, 1]) * stre / scalestr
# stretch[:,0] = stretchori[:,0]
stretch[:, 1] = stretchori[:, 1]
# stretch[:,2] = stretchori[:,2]
stretch[:, 3] = stretchori[:, 3]
print(stretch)
bend = bend
matconfig['density'] = density.detach().numpy().tolist()
matconfig['stretching'] = stretch.detach().numpy().tolist()
matconfig['bending'] = bend.detach().numpy().tolist()
save_config(matconfig, sys.argv[1] + '/mat.json')
return density, stretch, bend
def get_loss_eval(steps, sim):
fstd = sys.argv[2] + '/%04d_00.obj' % (steps / 2)
mesh = arcsim.Mesh()
arcsim.load_obj(mesh, fstd)
diffs = []
for node0, node1 in zip(mesh.nodes, sim.cloths[0].mesh.nodes):
diffs.append((node0.x - node1.x).norm())
arcsim.delete_mesh(mesh)
return torch.stack(diffs).mean()
def get_loss(steps,sim): fstd = sys.argv[2]+'/%04d_00.obj'%(steps/2) mesh = arcsim.Mesh() arcsim.load_obj(mesh, fstd) diffs = [] for node0,node1 in zip(mesh.nodes, sim.cloths[0].mesh.nodes): diffs.append((np.array(arcsim.tovec(node0.x))-np.array(arcsim.tovec(node1.x)))**2) arcsim.delete_mesh(mesh) return np.stack(diffs).sum(axis=1).mean()
def renew_loss():
print('renew',steps)
matconfig['density'] = density.detach().numpy().tolist()
matconfig['stretching'] = stretch.detach().numpy().tolist()
matconfig['bending'] = bend.detach().numpy().tolist()
save_config(matconfig, sys.argv[1]+'/mat.json')
arcsim.delete_mesh(sim.cloths[0].mesh)
reset_sim(sim)
loss = run_sim(steps, sim)
get_grad(steps, sim, loss, density,stretch,bend)
return loss
def do_train(cur_step, sim):
while True:
steps = 4 * 25 * spf
reset_sim(sim)
loss, dis = run_sim(steps, sim)
# loss = get_loss(steps,sim)
print('step={}'.format(cur_step))
print('loss={}'.format(loss.data))
f.write('step {}: loss={}\n'.format(cur_step, loss.data))
print('step {}: loss={}\n'.format(cur_step, loss.data))
#f.write('{}\n'.format(vext.grad));
arcsim.delete_mesh(sim.cloths[0].mesh)
break
def do_train(cur_step, optimizer, sim):
while True:
global steps
steps = min(50, cur_step * 2 + 2)
density, stretch, bend = reset_sim(sim)
st = time.time()
loss = run_sim(steps, sim)
en0 = time.time()
# loss = get_loss(steps,sim)
optimizer.zero_grad()
print('step={}'.format(cur_step))
print('forward time={}'.format(en0 - st))
print('loss={}'.format(loss.data))
f.write('step {}: d={} loss={}\n'.format(cur_step,
density.data * scaleden,
loss.data))
print('step {}: d={} loss={}\n'.format(cur_step,
density.data * scaleden,
loss.data))
if loss < 1e-4:
break
#loss.backward()
# dgrad, stgrad, begrad = torch.autograd.grad(loss, [density, stretch, bend])
#density.grad.data.clamp_(-1,1)
#stretch.grad.data.clamp_(-1,1)
#bend.grad.data.clamp_(-1e-3,1e-3)
#print(density.grad)
#print(stretch.grad)
#print(bend.grad)
optimizer.step(renew_loss)
en1 = time.time()
print('backward time={}'.format((en1 - st) / steps))
print(density * scaleden)
print(stretch * scalestr)
print(bend * scaleben)
ld = (torch.tensor(matstd['density'], dtype=torch.float64) -
density * scaleden).norm().data
ls = (torch.tensor(matstd['stretching'], dtype=torch.float64) -
stretch * scalestr).norm().data
lb = (torch.tensor(matstd['bending'], dtype=torch.float64) -
bend * scaleben).norm().data
print('dis={} {} {}\n'.format(ld, ls, lb))
matconfig['density'] = density.detach().numpy().tolist()
matconfig['stretching'] = stretch.detach().numpy().tolist()
matconfig['bending'] = bend.detach().numpy().tolist()
save_config(matconfig, sys.argv[1] + '/mat.json')
arcsim.delete_mesh(sim.cloths[0].mesh)
get_dis(density, stretch, bend)
