def infer_batch(self, image, label):
image = image.cuda()
label = label.cuda()
logit = self.net(image)
pred = logit.max(dim=1)[1]
return fast_hist(label, pred)
def inf_batch(self, image, label):
image = image #.cuda()
label = label #.cuda()
with torch.no_grad():
logit = self.net(image)
pred = logit.max(dim=1)[1]
self.hist += fast_hist(label, pred)
def eval(args):
torch.cuda.set_device(args.local_rank)
dist.init_process_group(backend='nccl',
init_method='tcp://127.0.0.1:{}'.format(
config_CS.port),
world_size=torch.cuda.device_count(),
rank=0)
dataset = CityScapes(mode='val')
sampler = torch.utils.data.distributed.DistributedSampler(dataset)
dataloader = DataLoader(dataset,
batch_size=1,
shuffle=False,
sampler=sampler,
num_workers=4,
pin_memory=True,
drop_last=False)
# net = Origin_Res()
# net = Deeplab_v3plus()
net = HighOrder(19)
net.cuda()
net = torch.nn.SyncBatchNorm.convert_sync_batchnorm(net)
net = nn.parallel.DistributedDataParallel(net,
device_ids=[args.local_rank],
output_device=args.local_rank)
net.load_state_dict(torch.load('./Res60000.pth'))
net.eval()
data = iter(dataloader)
num = 0
hist = 0
with torch.no_grad():
while 1:
try:
image, label, name = next(data)
except:
break
image = image.cuda()
label = label.cuda()
label = torch.squeeze(label, 1)
output = net(image)
pred = output.max(dim=1)[1]
hist_once = fast_hist(label, pred)
hist = torch.tensor(hist).cuda()
hist = hist + hist_once
dist.all_reduce(hist, dist.ReduceOp.SUM)
num += 1
if num % 50 == 0:
print('iter :{}'.format(num))
hist = hist.cpu().numpy().astype(np.float32)
miou = cal_scores(hist)
print('miou = {}'.format(miou))
def val_segmentation(epoch, net_segmentation):
global best_iou
global val_seg_iou
progbar = tqdm(total=len(val_seg_loader), desc='Val')
net_segmentation.eval()
val_seg_loss.append(0)
hist = np.zeros((nClasses, nClasses))
for batch_idx, (inputs_, targets) in enumerate(val_seg_loader):
inputs_, targets = Variable(inputs_.to(device)), Variable(
targets.to(device))
outputs = net_segmentation(inputs_)
total_loss = 1 - soft_iou(outputs, targets, ignore=ignore_class)
val_seg_loss[-1] += total_loss.data
_, predicted = torch.max(outputs.data, 1)
correctLabel = targets.view(-1, targets.size()[1], targets.size()[2])
hist += fast_hist(
correctLabel.view(correctLabel.size(0), -1).cpu().numpy(),
predicted.view(predicted.size(0), -1).cpu().numpy(), nClasses)
miou, p_acc, fwacc = performMetrics(hist)
progbar.set_description('Val (loss=%.4f, mIoU=%.4f)' %
(val_seg_loss[-1] / (batch_idx + 1), miou))
progbar.update(1)
val_seg_loss[-1] = val_seg_loss[-1] / len(val_seg_loader)
val_miou, _, _ = performMetrics(hist)
val_seg_iou += [val_miou]
if best_iou < val_miou:
best_iou = val_miou
print('Saving..')
state = {'net_segmentation': net_segmentation}
torch.save(state,
model_root + experiment + 'segmentation' + '.ckpt.t7')
def train_segmentation(epoch, net_segmentation, seg_optimizer):
global train_seg_iou
progbar = tqdm(total=len(train_seg_loader), desc='Train')
net_segmentation.train()
train_seg_loss.append(0)
seg_optimizer.zero_grad()
hist = np.zeros((nClasses, nClasses))
for batch_idx, (inputs_, targets) in enumerate(train_seg_loader):
inputs_, targets = Variable(inputs_.to(device)), Variable(
targets.to(device))
outputs = net_segmentation(inputs_)
total_loss = (
1 - soft_iou(outputs, targets, ignore=ignore_class)) / ITER_SIZE
total_loss.backward()
if (batch_idx % ITER_SIZE == 0
and batch_idx != 0) or batch_idx == len(train_loader) - 1:
seg_optimizer.step()
seg_optimizer.zero_grad()
train_seg_loss[-1] += total_loss.data
_, predicted = torch.max(outputs.data, 1)
correctLabel = targets.view(-1, targets.size()[1], targets.size()[2])
hist += fast_hist(
correctLabel.view(correctLabel.size(0), -1).cpu().numpy(),
predicted.view(predicted.size(0), -1).cpu().numpy(), nClasses)
miou, p_acc, fwacc = performMetrics(hist)
progbar.set_description('Train (loss=%.4f, mIoU=%.4f)' %
(train_seg_loss[-1] / (batch_idx + 1), miou))
progbar.update(1)
train_seg_loss[-1] = train_seg_loss[-1] / len(train_seg_loader)
miou, p_acc, fwacc = performMetrics(hist)
train_seg_iou += [miou]
def evaluate_segmentation(net_segmentation):
net_segmentation.eval()
hist = np.zeros((nClasses, nClasses))
val_seg_loader = torch.utils.data.DataLoader(segmentation_data_loader(
img_root=val_img_root,
gt_root=val_gt_root,
image_list=val_image_list,
suffix=dataset,
out=out,
crop=False,
mirror=False),
batch_size=1,
num_workers=8,
shuffle=False)
progbar = tqdm(total=len(val_seg_loader), desc='Eval')
hist = np.zeros((nClasses, nClasses))
for batch_idx, (inputs_, targets) in enumerate(val_seg_loader):
inputs_, targets = Variable(inputs_.to(device)), Variable(
targets.to(device))
outputs = net_segmentation(inputs_)
_, predicted = torch.max(outputs.data, 1)
correctLabel = targets.view(-1, targets.size()[1], targets.size()[2])
hist += fast_hist(
correctLabel.view(correctLabel.size(0), -1).cpu().numpy(),
predicted.view(predicted.size(0), -1).cpu().numpy(), nClasses)
miou, p_acc, fwacc = performMetrics(hist)
progbar.set_description('Eval (mIoU=%.4f)' % (miou))
progbar.update(1)
miou, p_acc, fwacc = performMetrics(hist)
print('\n mIoU: ', miou)
print('\n Pixel accuracy: ', p_acc)
print('\n Frequency Weighted Pixel accuracy: ', fwacc)
def eval(args):
torch.cuda.set_device(args.local_rank)
dist.init_process_group(backend='nccl',
init_method='tcp://127.0.0.1:{}'.format(
config.port),
world_size=torch.cuda.device_count(),
rank=args.local_rank
# rank=args.local_rank
)
dataset = ADE20K(mode='val')
sampler = torch.utils.data.distributed.DistributedSampler(dataset)
dataloader = DataLoader(dataset,
batch_size=1,
shuffle=False,
sampler=sampler,
num_workers=4,
drop_last=False,
pin_memory=True)
net = PANet(150)
net.cuda()
net = torch.nn.SyncBatchNorm.convert_sync_batchnorm(net)
net = nn.parallel.DistributedDataParallel(net,
device_ids=[args.local_rank],
output_device=args.local_rank)
net.load_state_dict(
torch.load('./GPADE20Kres50150000.pth', map_location='cpu'))
net.eval()
data = iter(dataloader)
palette = get_palette(256)
num = 0
hist = 0
with torch.no_grad():
while 1:
try:
image, label, name = next(data)
except:
break
image = image.cuda()
label = label.cuda()
label = torch.squeeze(label, 1)
N, _, H, W = image.size()
preds = torch.zeros((N, 150, H, W))
preds = preds.cuda()
for scale in config.eval_scales:
new_hw = [int(H * scale), int(W * scale)]
image_change = F.interpolate(image,
new_hw,
mode='bilinear',
align_corners=True)
output, w = net(image_change)
output = F.interpolate(output, (H, W),
mode='bilinear',
align_corners=True)
output = F.softmax(output, 1)
preds += output
if config.eval_flip:
output, w = net(torch.flip(image_change, dims=(3, )))
output = torch.flip(output, dims=(3, ))
output = F.interpolate(output, (H, W),
mode='bilinear',
align_corners=True)
output = F.softmax(output, 1)
preds += output
pred = preds.max(dim=1)[1]
hist_once = fast_hist(label, pred)
hist = torch.tensor(hist).cuda()
hist = hist + hist_once
dist.all_reduce(hist, dist.ReduceOp.SUM)
num += 1
if num % 5 == 0:
print('iter: {}'.format(num))
preds = np.asarray(np.argmax(preds.cpu(), axis=1), dtype=np.uint8)
# for i in range(preds.shape[0]):
# pred = convert_label(preds[i], inverse=True)
# save_img = Image.fromarray(pred)
# save_img.putpalette(palette)
# save_img.save(os.path.join('./CS_results/', name[i] + '.png'))
hist = hist.cpu().numpy().astype(np.float32)
miou = cal_scores(hist)
print('miou = {}'.format(miou))
