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 main(ckp_name='final.pth'):
sess = Session(dt_split='val')
sess.load_checkpoints(ckp_name)
dt_iter = sess.dataloader
sess.net.eval()
for i, [image, label] in enumerate(dt_iter):
sess.inf_batch(image, label)
if i % 10 == 0:
logger.info('num-%d' % i)
scores, cls_iu = cal_scores(sess.hist.cpu().numpy())
for k, v in scores.items():
logger.info('%s-%f' % (k, v))
scores, cls_iu = cal_scores(sess.hist.cpu().numpy())
for k, v in scores.items():
logger.info('%s-%f' % (k, v))
logger.info('')
for k, v in cls_iu.items():
logger.info('%s-%f' % (k, v))
def eval_epoch(writer, iterations):
sess = Session(dt_split='val')
sess.load_checkpoints('latest.pth')
dt_iter = sess.dataloader
sess.net.eval()
for i, [image, label] in enumerate(dt_iter):
sess.inf_batch(image, label)
if i % 10 == 0:
logger.info('num-%d' % i)
scores, cls_iu = cal_scores(sess.hist.cpu().numpy())
for k, v in scores.items():
logger.info('%s-%f' % (k, v))
scores, cls_iu = cal_scores(sess.hist.cpu().numpy())
for k, v in scores.items():
logger.info('%s-%f' % (k, v))
logger.info('')
for k, v in cls_iu.items():
logger.info('%s-%f' % (k, v))
writer.add_scalar('mIoU', scores['mIoU'], iterations)
writer.add_scalar('Acc', scores['pAcc'], iterations)
def eval_by_steps(ckp_name):
sess = EvalSession('val')
sess.load_checkpoints(ckp_name)
dt_iter = sess.dataloader
sess.net.eval()
for _, [image, label] in enumerate(dt_iter):
sess.inf_batch(image, label)
logger.info('')
logger.info('Total Scores')
scores, cls_iu = cal_scores(sess.hist.cpu().numpy())
for k, v in scores.items():
logger.info('%s-%f' % (k, v))
logger.info('')
logger.info('Class Scores')
for k, v in cls_iu.items():
logger.info('%s-%f' % (k, v))
return scores['mIoU']
def val(self, prefix='', logging=False, steps=None):
self.net.eval()
hist = 0
for image, label in self.val_loader:
hist += self.infer_batch(image, label)
scores, cls_iu = cal_scores(hist.cpu().numpy())
mIoU = scores['mIoU']
if mIoU > self.best_mIoU:
self.best_mIoU = mIoU
if self.split == 'trainval':
path = 'best.pth'
else:
path = 'best_trainaug.pth'
self.save_checkpoints(path)
steps = self.step if steps is None else steps
for k, v in scores.items():
self.writer.add_scalar(prefix + k, v, steps)
if logging:
logger.info('')
logger.info('Total Scores')
for k, v in scores.items():
logger.info('%s-%f' % (k, v))
logger.info('')
logger.info('Class Scores')
for k, v in cls_iu.items():
logger.info('%s-%f' % (k, v))
else:
self.net.train()
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))