if cuda:
net.cuda()
if num_GPU > 1:
net = nn.DataParallel(net)
if pre_trained:
net.load_state_dict(torch.load('%s/model/' % out_file + path + 'netG.pth'))
# print('Load success!')
else:
pass
# net.apply(weights_init)
########### LOSS & OPTIMIZER ##########
criterion = nn.CrossEntropyLoss(ignore_index=255)
optimizer = torch.optim.Adam(net.parameters(), lr=learning_rate, betas=(beta1, 0.999))
metric = SegmentationMetric(4)
early_stopping = EarlyStopping(patience=7, verbose=True)
if __name__ == '__main__':
start = time.time()
net.train()
for epoch in range(1, niter+1):
lr_adjust = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, 10, eta_min=0.0, last_epoch=-1)
for i in range(0, train_datatset_.__len__(), batch_size):
train_datatset_next = train_datatset_.__next__()
train_loader = torch.utils.data.DataLoader(dataset=train_datatset_next, batch_size=batch_size, shuffle=True,
num_workers=num_workers)
for initial_image, semantic_image in train_loader:
# print(initial_image.shape)
initial_image = initial_image.cuda()
semantic_image = semantic_image.cuda()
cuda = True
path = '2017'
if path == '2017':
time_series = 7
else:
time_series = 4
data_path = './' + path + 'data/test'
test_datatset_ = train_dataset(data_path, time_series=time_series)
model = MSFCN2D(time_series, 4, 4)
model_path = './checkpoint/' + model.name + '/model/' + path + 'netG.pth'
model.load_state_dict(torch.load(model_path, map_location='cpu'))
model.eval()
metric = SegmentationMetric(4)
out_file = './result/' + model.name + '/' + path + '/'
if cuda:
model.cuda()
try:
os.makedirs(out_file)
except OSError:
pass
if __name__ == '__main__':
start = time.time()
for i in range(0, test_datatset_.__len__(), 1):
test_datatset_next = test_datatset_.__next__()
test_loader = torch.utils.data.DataLoader(dataset=test_datatset_next,
try:
os.makedirs(out_file)
os.makedirs(out_file + '/')
except OSError:
pass
if cuda:
net.cuda()
if num_GPU > 1:
net = nn.DataParallel(net)
########### LOSS & OPTIMIZER ##########
criterion = nn.CrossEntropyLoss(ignore_index=255)
optimizer = torch.optim.Adam(net.parameters(), lr=learning_rate)
metric = SegmentationMetric(class_num)
early_stopping = EarlyStopping(patience=10, verbose=True)
if __name__ == '__main__':
start = time.time()
net.train()
lr_adjust = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, 10, eta_min=learning_rate * 0.01, last_epoch=-1)
for epoch in range(1, niter + 1):
for iter_num in trange(2000 // index, desc='train, epoch:%s' % epoch):
train_iter = train_datatset_.data_iter_index(index=index)
for initial_image, semantic_image in train_iter:
# print(initial_image.shape)
initial_image = initial_image.cuda()
semantic_image = semantic_image.cuda()