split='test',
data_augmentation=False)
idx = opt.idx
print("model %d/%d" % (idx, len(d)))
point, seg = d[idx]
print(point.size(), seg.size())
point_np = point.numpy()
cmap = plt.cm.get_cmap("hsv", 10)
cmap = np.array([cmap(i) for i in range(10)])[:, :3]
gt = cmap[seg.numpy() - 1, :]
state_dict = torch.load(opt.model, map_location=torch.device('cpu'))
classifier = PointNetDenseCls(k=state_dict['conv4.weight'].size()[0])
classifier.load_state_dict(state_dict)
classifier.eval()
point = point.transpose(1, 0).contiguous()
point = Variable(point.view(1, point.size()[0], point.size()[1]))
pred, _, _ = classifier(point)
pred_choice = pred.data.max(2)[1]
print(pred_choice)
#print(pred_choice.size())
pred_color = cmap[pred_choice.numpy()[0], :]
#print(pred_color.shape)
showpoints(point_np, gt, pred_color)
batch_size=opt.batchSize,
shuffle=True,
num_workers=int(opt.workers))
print(len(dataset), len(test_dataset)) #打印数据集和测试集的对象个数
num_classes = dataset.num_seg_classes #从misc/num_seg_classes.txt读取分割个数
print('classes', num_classes) #例如chair的分割个数是4
try:
os.makedirs(opt.outf) #用于递归创建目录
except OSError:
pass
blue = lambda x: '\033[94m' + x + '\033[0m' #在训练时，将test设置成蓝色字底
classifier = PointNetDenseCls(
k=num_classes,
feature_transform=opt.feature_transform) #读取model.py的densecls函数
if opt.model != '':
classifier.load_state_dict(torch.load(opt.model)) #如果有预训练模型，加载预训练模型
optimizer = optim.Adam(classifier.parameters(), lr=0.001,
betas=(0.9, 0.999)) #优化器
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=20, gamma=0.5)
classifier.cuda()
num_batch = len(dataset) / opt.batchSize #数据集分成batch的个数
for epoch in range(opt.nepoch): #执行一个epoch
scheduler.step()
for i, data in enumerate(dataloader, 0): #enumerate枚举
testdataloader = torch.utils.data.DataLoader(test_dataset,
batch_size=opt.batchSize,
shuffle=True,
num_workers=int(opt.workers))
print(len(dataset), len(test_dataset))
num_classes = dataset.num_seg_classes
print('classes', num_classes)
try:
os.makedirs(opt.outf)
except OSError:
pass
blue = lambda x: '\033[94m' + x + '\033[0m'
classifier = PointNetDenseCls(k=num_classes,
feature_transform=opt.feature_transform)
if opt.model != '':
classifier.load_state_dict(torch.load(opt.model))
optimizer = optim.Adam(classifier.parameters(), lr=0.001, betas=(0.9, 0.999))
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=20, gamma=0.5)
classifier.cuda()
num_batch = len(dataset) / opt.batchSize
for epoch in range(opt.nepoch):
scheduler.step()
for i, data in enumerate(dataloader, 0):
points, target = data
points = points.transpose(2, 1)
testdataloader = torch.utils.data.DataLoader(test_dataset,
batch_size=opt.batchsize,
shuffle=True,
num_workers=int(opt.workers))
print(len(dataset), len(test_dataset))
try:
os.makedirs(opt.outf)
except OSError:
pass
blue = lambda x: '\033[94m' + x + '\033[0m'
device = opt.device
netG = PointNetDenseCls(device=device, feature_transform=opt.feature_transform)
localD = LocalDiscriminator(k=2, device=device)
globalD = GlobalDiscriminator(k=2, device=device)
if opt.model != '':
netG.load_state_dict(torch.load(opt.model))
optimizerG = optim.Adam(netG.parameters(), lr=0.001, betas=(0.9, 0.999))
optimizerD = optim.Adam(list(globalD.parameters()) + list(localD.parameters()),
lr=0.0005,
betas=(0.9, 0.999))
schedulerG = optim.lr_scheduler.StepLR(optimizerG, step_size=20, gamma=0.5)
schedulerD = optim.lr_scheduler.StepLR(optimizerD, step_size=20, gamma=0.5)
netG.to(device)
def train(lr=0.001):
parser = argparse.ArgumentParser()
opt = parser.parse_args()
opt.nepoch = 1
opt.batchsize = 18
opt.workers = 0
opt.outf = 'completion'
opt.dataset = '/home/cdi0/data/shape_net_core_uniform_samples_2048_split/'
opt.feature_transform = False
opt.model = ''
opt.device = 'cuda:1'
opt.lr = lr
opt.manualSeed = random.randint(1, 10000) # fix seed
print("Random Seed: ", opt.manualSeed)
random.seed(opt.manualSeed)
torch.manual_seed(opt.manualSeed)
dataset = ShapeNetDataset(dir=opt.dataset, )
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=opt.batchsize,
shuffle=True,
num_workers=int(opt.workers))
test_dataset = ShapeNetDataset(
dir=opt.dataset,
train='test',
)
testdataloader = torch.utils.data.DataLoader(test_dataset,
batch_size=opt.batchsize,
shuffle=True,
num_workers=int(opt.workers))
print(len(dataset), len(test_dataset))
#try:
# os.makedirs(opt.outf)
#except OSError:
# pass
blue = lambda x: '\033[94m' + x + '\033[0m'
device = opt.device
netG = PointNetDenseCls(device=device,
feature_transform=opt.feature_transform)
localD = LocalDiscriminator(k=2, device=device)
globalD = GlobalDiscriminator(k=2, device=device)
if opt.model != '':
netG.load_state_dict(torch.load(opt.model))
optimizerG = optim.Adam(netG.parameters(), lr=0.001, betas=(0.9, 0.999))
optimizerD = optim.Adam(list(globalD.parameters()) +
list(localD.parameters()),
lr=0.001,
betas=(0.9, 0.999))
schedulerG = optim.lr_scheduler.StepLR(optimizerG, step_size=20, gamma=0.5)
schedulerD = optim.lr_scheduler.StepLR(optimizerD, step_size=20, gamma=0.5)
netG.to(device)
localD.to(device)
globalD.to(device)
criterion = distChamfer
Dcriterion = nn.BCELoss()
#Dcriterion = F.nll_loss
real_label = 1
fake_label = 0
num_batch = len(dataset) / opt.batchsize
writer = SummaryWriter()
for epoch in range(opt.nepoch):
for i, data in (enumerate(dataloader, 0)):
#k = 614
points, target, mask = data # Nx4 or Nx3
points = points.transpose(2, 1) # 4xN
points, target = points.to(device, dtype=torch.float), target.to(
device, dtype=torch.float)
b_size = points.shape[0]
mask_ = mask.unsqueeze(2).repeat(1, 1, 3)
#print(mask_.any(dim = 2).sum(dim=1))
mask__ = ~mask_
#print(mask__.any(dim = 2).sum(dim=1))
mask__ = mask__.to(device, dtype=torch.float32)
mask_ = mask_.to(device, dtype=torch.float32)
optimizerD.zero_grad()
localD = localD.train()
globalD = globalD.train()
###### train D ######
#label_real = torch.stack((torch.zeros(b_size),torch.ones(b_size)), dim = 1).to(device, dtype = torch.long)
#label_fake = torch.stack((torch.ones(b_size),torch.zeros(b_size)), dim = 1).to(device, dtype = torch.long)
label = torch.full((b_size, ), real_label, device=device)
#print(mask__)
#print(mask__[mask__.sum(dim=2) != 0].shape)
target_mask = mask__ * target
target_mask = target_mask[torch.abs(target_mask).sum(
dim=2) != 0].view(b_size, -1, 3)
target, target_mask = target.transpose(
2, 1).contiguous(), target_mask.transpose(2, 1).contiguous()
output_g = globalD(target)
output_l = localD(target_mask)
#rint(output_g.shape)
#rint(output_l.shape)
#rint(label.shape)
errD_real_g = Dcriterion(output_g, label)
errD_real_l = Dcriterion(output_l, label)
errD_real = errD_real_g + errD_real_l
errD_real.backward()
target = target.transpose(2, 1).contiguous()
pred = netG(points)
#rint(pred.shape)
##int(target.shape)
#rint(mask_.shape)
#rint(mask__.shape)
pred = (pred * mask__) + (target * mask_)
pred_mask = pred * mask__
pred_mask = pred_mask[torch.abs(pred_mask).sum(dim=2) != 0].view(
b_size, -1, 3)
pred, pred_mask = pred.transpose(
2, 1).contiguous(), pred_mask.transpose(2, 1).contiguous()
output_g = globalD(pred.detach())
output_l = localD(pred_mask.detach())
label.fill_(fake_label)
errD_fake_g = Dcriterion(output_g, label)
errD_fake_l = Dcriterion(output_l, label)
errD_fake = errD_fake_g + errD_fake_l
errD_fake.backward()
errD = errD_real + errD_fake
if errD.item() > 0.1:
optimizerD.step()
###### train G ######
optimizerG.zero_grad()
optimizerD.zero_grad()
netG = netG.train()
output_g = globalD(pred)
output_l = localD(pred_mask)
label.fill_(real_label)
errG_g = Dcriterion(output_g, label)
errG_l = Dcriterion(output_l, label)
errG = errG_g + errG_l
pred = pred.transpose(2, 1).contiguous()
#rint(pred.shape)
#rint(target.shape)
dist1, dist2 = criterion(pred, target)
chamferloss = (torch.mean(dist1)) + (torch.mean(dist2))
loss = chamferloss + errG
loss.backward()
if opt.feature_transform:
loss += feature_transform_regularizer(trans_feat) * 0.001
optimizerG.step()
print('[%d: %d/%d] D_loss: %f, G_loss: %f, Chamfer_loss: %f ' %
(epoch, i, num_batch, errD.item(), errG.item(),
chamferloss.item()))
if i % 10 == 0:
j, data = next(enumerate(testdataloader, 0))
points, target, mask = data
points = points.transpose(2, 1)
points, target = points.to(
device, dtype=torch.float), target.to(device,
dtype=torch.float)
b_size = points.shape[0]
localD = localD.eval()
globalD = globalD.eval()
###### eval D ######
label = torch.full((b_size, ), real_label, device=device)
#label_real = torch.stack((torch.zeros(b_size),torch.ones(b_size)), dim = 1).to(device)
#label_fake = torch.stack((torch.ones(b_size),torch.zeros(b_size)), dim = 1).to(device)
mask_ = mask.unsqueeze(2).repeat(1, 1, 3)
mask__ = ~mask_
mask__ = mask__.to(device, dtype=torch.float32)
mask_ = mask_.to(device, dtype=torch.float32)
target_mask = mask__ * target
target_mask = target_mask[torch.abs(target_mask).sum(
dim=2) != 0].view(b_size, -1, 3)
target, target_mask = target.transpose(
2, 1).contiguous(), target_mask.transpose(2,
1).contiguous()
output_g = globalD(target)
output_l = localD(target_mask)
errD_real_g_eval = Dcriterion(output_g, label)
errD_real_l_eval = Dcriterion(output_l, label)
errD_real_eval = errD_real_g_eval + errD_real_l_eval
target = target.transpose(2, 1).contiguous()
pred = netG(points)
pred = (pred * mask__) + (target * mask_)
pred_mask = pred * mask__
pred_mask = pred_mask[torch.abs(pred_mask).sum(
dim=2) != 0].view(b_size, -1, 3)
pred, pred_mask = pred.transpose(
2, 1).contiguous(), pred_mask.transpose(2, 1).contiguous()
output_g_eval = globalD(pred.detach())
output_l_eval = localD(pred_mask.detach())
label.fill_(fake_label)
errD_fake_g_eval = Dcriterion(output_g, label)
errD_fake_l_eval = Dcriterion(output_l, label)
errD_fake_eval = errD_fake_g_eval + errD_fake_l_eval
errD_eval = errD_real_eval + errD_fake_eval
###### eval G ######
netG = netG.eval()
output_g = globalD(pred)
output_l = localD(pred_mask)
label.fill_(real_label)
errG_g_eval = Dcriterion(output_g, label)
errG_l_eval = Dcriterion(output_l, label)
errG_eval = errG_g_eval + errG_l_eval
pred = pred.transpose(2, 1).contiguous()
dist1, dist2 = criterion(pred, target)
chamferloss_eval = (torch.mean(dist1)) + (torch.mean(dist2))
loss_eval = chamferloss_eval + errG_eval
print('[%d: %d/%d] %s D_loss: %f, G_loss: %f ' %
(epoch, i, num_batch, blue('test'), errD_eval.item(),
loss.item()))
if i % 100 == 0:
n = int(i / 100)
writer.add_scalar('errD_real', errD_real.item(),
27 * epoch + n)
writer.add_scalar('errD_fake', errD_fake.item(),
27 * epoch + n)
writer.add_scalar('errD_loss', errD.item(), 27 * epoch + n)
writer.add_scalar('validation errD_real',
errD_real_eval.item(), 27 * epoch + n)
writer.add_scalar('validation errD_fake',
errD_fake_eval.item(), 27 * epoch + n)
writer.add_scalar('validation errD_loss', errD_eval.item(),
27 * epoch + n)
writer.add_scalar('errG_global', errG_g.item(), 27 * epoch + n)
writer.add_scalar('errG_local', errG_l.item(), 27 * epoch + n)
writer.add_scalar('chamfer_loss', chamferloss.item(),
27 * epoch + n)
writer.add_scalar('errG_loss', loss.item(), 27 * epoch + n)
writer.add_scalar('validation errG_global', errG_g_eval.item(),
27 * epoch + n)
writer.add_scalar('validation errG_local', errG_l_eval.item(),
27 * epoch + n)
writer.add_scalar('validation chamfer_loss',
chamferloss_eval.item(), 27 * epoch + n)
writer.add_scalar('validation errG_loss', loss_eval.item(),
27 * epoch + n)
for name, param in globalD.named_parameters():
writer.add_histogram(name,
param.clone().cpu().data.numpy(),
27 * epoch + n)
for name, param in localD.named_parameters():
writer.add_histogram(name,
param.clone().cpu().data.numpy(),
27 * epoch + n)
for name, param in netG.named_parameters():
writer.add_histogram(name,
param.clone().cpu().data.numpy(),
27 * epoch + n)
schedulerG.step()
schedulerD.step()
#torch.save(netG.state_dict(), '%s/com_model_G_%f_%d.pth' % (opt.outf, loss.item(), epoch))
#torch.save(localD.state_dict(), '%s/com_model_localD_%f_%d.pth' % (opt.outf, errD.item(), epoch))
#torch.save(globalD.state_dict(), '%s/com_model_globalD_%f_%d.pth' % (opt.outf, errD.item(), epoch))
return errD.item(), errG, chamferloss
print(opt)
d = ShapeNetDataset(
dir=opt.dataset,
train='test',
)
device = opt.device
idx = opt.idx
print(d.lst[idx])
print("model %d/%d" % (idx, len(d)))
point, target, mask = d[idx]
print(point.shape, target.shape)
#point_np = point.numpy()
state_dict = torch.load(opt.model, map_location='cpu')
classifier = PointNetDenseCls(device=device)
classifier.load_state_dict(state_dict)
classifier.to(device)
classifier.eval()
input_cloud = PyntCloud(
pd.DataFrame(
# same arguments that you are passing to visualize_pcl
data=point[:, :3],
columns=["x", "y", "z"]))
input_cloud.to_file("input.ply")
target_cloud = PyntCloud(
pd.DataFrame(
# same arguments that you are passing to visualize_pcl
data=target,
batch_size=opt.batchSize,
shuffle=True,
num_workers=int(opt.workers),
drop_last=True)
print(len(dataset), len(test_dataset))
num_classes = dataset.num_seg_classes
print('classes', num_classes)
try:
os.makedirs(opt.outf)
except OSError:
pass
blue = lambda x: '\033[94m' + x + '\033[0m'
classifier = PointNetDenseCls(k=num_classes,
feature_transform=opt.feature_transform)
if opt.gpu != -1:
classifier = torch.nn.DataParallel(classifier).to(device)
else:
classifier.to(device)
if opt.model != '':
classifier.load_state_dict(torch.load(opt.model))
optimizer = optim.Adam(classifier.parameters(),
lr=0.001,
betas=(0.9, 0.999))
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=20, gamma=0.5)
num_batch = len(dataset) / opt.batchSize
data_augmentation=False)
testdataloader = torch.utils.data.DataLoader(
test_dataset,
batch_size=opt.batchSize,
shuffle=True,
num_workers=int(opt.workers))
print(len(dataset), len(test_dataset))
num_classes = dataset.num_seg_classes
print('classes', num_classes)
try:
os.makedirs(opt.outf)
except OSError:
pass
classifier = PointNetDenseCls(k=num_classes, feature_transform=opt.feature_transform)
if opt.model != '':
classifier.load_state_dict(torch.load(opt.model))
optimizer = optim.Adam(classifier.parameters(), lr=0.001, betas=(0.9, 0.999))
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=20, gamma=0.5)
classifier.cuda()
num_batch = len(dataset) / opt.batchSize
def get_iou(pred_np, target_np):
shape_ious = []
for shape_idx in range(target_np.shape[0]):
parts = range(num_classes) # np.unique(target_np[shape_idx])
dataset = ShapeNetDataset(root=dataset_path,
classification=False,
class_choice=[class_choice])
test_dataset = ShapeNetDataset(root=dataset_path,
classification=False,
class_choice=[class_choice],
split='test',
data_augmentation=False)
# print(len(dataset), len(test_dataset))
num_classes = dataset.num_seg_classes
print('segmentation classes', num_classes)
classifier = PointNetDenseCls(k=num_classes, feature_transform=False)
model_path = 'trained/seg/seg_model_' + 'Chair_' + str(model_num) + '.pth'
classifier.load_state_dict(torch.load(model_path))
classifier.cuda()
#%%
points, target = dataset[data_num]
points.unsqueeze_(0)
target.unsqueeze_(0)
points = points.transpose(2, 1)
points, target = points.cuda(), target.cuda()
classifier = classifier.eval()
pred, _, _ = classifier(points)
pred_choice = pred.data.max(2)[1]
train=False,
test_area='Area_6')
idx = 100
print("model %d/%d" % (idx, len(data)))
point, seg = data[idx]
print(point.size(), seg.size())
point_np = point.numpy()
#forward
num = 13
print('num={}'.format(num))
classifier = PointNetDenseCls(k=num)
classifier.load_state_dict(torch.load(opt.model))
classifier.eval()
point = point.transpose(1, 0).contiguous()
point = Variable(point.view(1, point.size()[0], point.size()[1]))
pred, _ = classifier(point)
print(pred.size())
pred_choice = pred.data.max(2)[1]
print(pred_choice.size())
print(point.size())
# print(point[0, :, 0])
point = point.squeeze()
# print(point[:, 0])
