def evaluate():
## calculate the average IOU between pred and gt
if os.path.exists(resume):
checkoint = torch.load(resume)
start_epoch = checkoint['epoch']
model.load = model.load_state_dict(checkoint['model'])
#optimizer.load_state_dict(checkoint['optim'])
print('load the resume checkpoint from epoch{}'.format(start_epoch))
else:
print("no resume checkpoint to load")
model.eval()
IOUs = 0
total_correct = 0
data_eval = multiDataset(data_rootpath, test=True)
eval_loader = torch.utils.data.DataLoader(data_eval,
batch_size=args.batch_size,
shuffle=True,
collate_fn=multi_collate)
# model.SingleNet
if os.path.exists(singlemodel_path):
t1 = time.time()
if is_GPU:
checkoint = torch.load(singlemodel_path)
else:
checkoint = torch.load(singlemodel_path,
map_location=lambda storage, loc: storage)
## notice :here we should map the order of GPU
## when the SingleModel is trained in GPU-1,we should map the model to GPU-0
model.SingleNet.load = model.SingleNet.load_state_dict(
checkoint['model'])
t2 = time.time()
print('singleNetwork load resume model from epoch{} use {}s'.format(
checkoint['epoch'], t2 - t1))
if is_GPU:
model.SingleNet = model.SingleNet.cuda()
for batch_idx, (img1s, img2s, targets, _, v12s) in enumerate(eval_loader):
## when predict,the target is target1 (img1)
if is_GPU:
img1s = Variable(img1s.cuda(), volatile=True)
img2s = Variable(img2s.cuda(), volatile=True)
targets = Variable(targets.cuda())
else:
img1s = Variable(img1s, volatile=True)
img2s = Variable(img2s, volatile=True)
targets = Variable(targets)
t1 = time.time()
# print v12s
outputs = model.predict(img1s, img2s, v12s, iter=args.iter)
t2 = time.time()
print('in batch{}/{} use {}s'.format(batch_idx * args.batch_size,
len(eval_loader.dataset),
t2 - t1))
#occupy=(outputs.data>0.5)if is_GPU else (outputs.data>0.5).type(torch.FloatTensor)
#print occupy.sum()
occupy = (outputs.data > 0.5)
for idx, target in enumerate(targets):
#correct+=(occupy[idx].eq(target.data)).sum()
insect = (target.data[occupy[idx]]).sum()
union = target.data.sum() + occupy[idx].sum() - insect
iou = insect * 1.0 / union
IOUs += iou
total_correct += insect
#print 'correct num:{}'.format(total_correct)
print('the average correct rate:{}'.format(total_correct * 1.0 /
(len(eval_loader.dataset))))
print('the average iou:{}'.format(IOUs * 1.0 / (len(eval_loader.dataset))))
singlemodel_path = './single_model/csg_single_train_ce_server_87.pth'
#model=singleNet()
model = MulitUpdateNet_deeper()
if os.path.exists(resume):
if is_GPU:
model.load_state_dict(torch.load(resume)['model'])
else:
model.load_state_dict(
torch.load(resume,
map_location=lambda storage, loc: storage)['model'])
print('load trained model success')
data_rootpath = './dataset/CsgData'
dataset = multiDataset(data_rootpath, 'csg', test=True)
def generate_binvox(num1, num2, demo_path='./demo2/'):
if os.path.exists(singlemodel_path):
t1 = time.time()
if is_GPU:
checkoint = torch.load(singlemodel_path)
else:
checkoint = torch.load(singlemodel_path,
map_location=lambda storage, loc: storage)
## notice :here we should map the order of GPU
## when the SingleModel is trained in GPU-1,we should map the model to GPU-0
model.SingleNet.load = model.SingleNet.load_state_dict(
checkoint['model'])
'''
a=np.random.randint(0,2,(2,64,64,64))
a=Variable(torch.from_numpy(a).type(torch.FloatTensor))
if is_GPU:
a = Variable(torch.from_numpy(a).cuda().type(torch.FloatTensor))
b= dense2sparse(a)
print (b[0].data.size()) ##list
a=sparse2dense(b)
print a
'''
data_rootpath = '../dataset/CubeData'
dataset = multiDataset(data_rootpath)
img_id, test_img = dataset.pull_img(10, 2)
test_img = np.array(test_img)
#print test_img
#cv2.imshow('mat',test_img)
#cv2.waitKey(2000)
#print img_id ## img read bingo
'''
img1,img2,target1,v12 = dataset.pull_item(10,2,7)
cv2.imshow('img1',img1.numpy())
cv2.imshow('img2',img2.numpy())
cv2.waitKey(2000)
print target1.numpy()
target_data= target1.numpy().astype(np.bool)