snapshot = 10000 paramName = 'models/'+exp_prefix+'stereo_2' predModel = 'models/9-3_stereo_2_100000.pkl' lossfilename = exp_prefix+'loss' SceneTurn = 5 ImgHeight = 320 ImgWidth = 640 stereonet = StereoNet() stereonet.cuda() loadPretrain(stereonet,predModel) normalize = Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225]) sceneDataset = SceneflowDataset(transform=Compose([ RandomCrop(size=(ImgHeight,ImgWidth)), RandomHSV((10,80,80)), ToTensor(), normalize])) kittiDataset = KittiDataset(transform=Compose([ RandomCrop(size=(ImgHeight,ImgWidth)), RandomHSV((7,50,50)), ToTensor(), normalize]), surfix='train') sceneDataloader = DataLoader(sceneDataset, batch_size=batch, shuffle=True, num_workers=4) kittiDataloader = DataLoader(kittiDataset, batch_size=batch, shuffle=True, num_workers=4) sceneiter = iter(sceneDataloader) kittiiter = iter(kittiDataloader) criterion = nn.SmoothL1Loss() # stereoOptimizer = optim.Adam(stereonet.parameters(), lr = Lr)
batch = 1
trainstep = 100000
showiter = 20
snapshot = 10000
paramName = 'models/' + exp_prefix + 'stereo_2'
predModel = 'models/9-2-2_stereo_2_50000.pkl'
lossfilename = exp_prefix + 'loss'
stereonet = StereoNet()
stereonet.cuda()
# loadPretrain(stereonet,predModel)
normalize = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
sceneDataset = SceneflowDataset(transform=Compose([
RandomCrop(size=(320, 640)),
RandomHSV((7, 37, 37)),
ToTensor(), normalize
]))
dataloader = DataLoader(sceneDataset,
batch_size=batch,
shuffle=True,
num_workers=8)
criterion = nn.SmoothL1Loss()
# stereoOptimizer = optim.Adam(stereonet.parameters(), lr = Lr)
stereoOptimizer = optim.Adam([{
'params': stereonet.preLoadedParams,
'lr': Lr
}, {
'params': stereonet.params
}],
predModel = 'models/1-4_stereo_2_45000.pkl'
dnetPreModel = 'models/4-2_dnet_2000.pkl'
stereonet = StereoNet()
stereonet.cuda()
loadPretrain(stereonet, predModel)
print '---'
dnet = DNet()
dnet.cuda()
loadPretrain(dnet, dnetPreModel)
normalize = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
sceneDataset = SceneflowDataset(transform=Compose([
RandomCrop(size=(320, 640)),
RandomHSV((10, 100, 100)),
ToTensor(), normalize
]))
dataloader = DataLoader(sceneDataset,
batch_size=batch,
shuffle=True,
num_workers=8)
criterion = nn.BCELoss()
# stereoOptimizer = optim.Adam(stereonet.parameters(), lr = Lr)
dnetOptimizer = optim.Adam(dnet.parameters(), lr=Lr)
label = torch.FloatTensor(batch)
input = torch.FloatTensor(batch, 320, 640)
real_label = 1
fake_label = 0
return gradient_penalty
stereonet = StereoNet()
stereonet.cuda()
loadPretrain(stereonet, predModel)
print '---'
dnet = DNet()
dnet.cuda()
# loadPretrain(dnet,dnetPreModel)
normalize = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
sceneDataset = SceneflowDataset(transform=Compose([
RandomCrop(size=(320, 640)),
RandomHSV((5, 30, 30)),
ToTensor(), normalize
]))
dataloader = DataLoader(sceneDataset,
batch_size=batch,
shuffle=True,
num_workers=8)
dataiter = iter(dataloader)
criterion1 = nn.SmoothL1Loss()
# stereoOptimizer = optim.Adam(stereonet.parameters(), lr = Lr)
# stereoOptimizer = optim.Adam([{'params':stereonet.preLoadedParams,'lr': 0},
# {'params':stereonet.params}], lr = stereo_lr)
# dnetOptimizer = optim.Adam(dnet.parameters(), lr = dnet_lr)
# paramName = 'models/'+exp_prefix+'stereo_2'
predModel = 'models/12-3-4_stereo_gan_80000.pkl'
dataset = 'scene'
stereonet = StereoNet()
stereonet.cuda()
loadPretrain(stereonet, predModel)
normalize = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
if dataset == 'scene':
sceneDataset = SceneflowDataset(filename='val.txt',
transform=Compose([
RandomCrop(size=(320, 640)),
RandomHSV((0, 0, 0)),
ToTensor(), normalize
]))
else:
sceneDataset = KittiDataset(transform=Compose([
RandomCrop(size=(320, 640)),
RandomHSV((0, 0, 0)),
ToTensor(), normalize
]))
dataloader = DataLoader(sceneDataset,
batch_size=1,
shuffle=True,
num_workers=8)
# criterion = nn.SmoothL1Loss()