def __init__(
self,
config,
device,
transform=transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize(
mean=StereoMatcherBase.IMAGENET_MEAN,
std=StereoMatcherBase.IMAGENET_STD,
),
]
),
):
StereoMatcherBase.__init__(self, config=config, device=device, transform=transform)
self._model = nets.AANet(
config.max_disp,
num_downsample=config.num_downsample,
feature_type=config.feature_type,
no_feature_mdconv=config.no_feature_mdconv,
feature_pyramid=config.feature_pyramid,
feature_pyramid_network=config.feature_pyramid_network,
feature_similarity=config.feature_similarity,
aggregation_type=config.aggregation_type,
num_scales=config.num_scales,
num_fusions=config.num_fusions,
num_stage_blocks=config.num_stage_blocks,
num_deform_blocks=config.num_deform_blocks,
no_intermediate_supervision=config.no_intermediate_supervision,
refinement_type=config.refinement_type,
mdconv_dilation=config.mdconv_dilation,
deformable_groups=config.deformable_groups,
)
self._model = self._model.to(device)
utils.load_pretrained_net(self._model, config.model_path, no_strict=True)
self._model.eval()
def main():
# For reproducibility
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
torch.backends.cudnn.benchmark = True
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Test loader
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)])
test_data = dataloader.StereoDataset(data_dir=args.data_dir,
dataset_name=args.dataset_name,
mode=args.mode,
save_filename=True,
transform=test_transform)
test_loader = DataLoader(dataset=test_data, batch_size=args.batch_size, shuffle=False,
num_workers=args.num_workers, pin_memory=True, drop_last=False)
aanet = nets.AANet(args.max_disp,
num_downsample=args.num_downsample,
feature_type=args.feature_type,
no_feature_mdconv=args.no_feature_mdconv,
feature_pyramid=args.feature_pyramid,
feature_pyramid_network=args.feature_pyramid_network,
feature_similarity=args.feature_similarity,
aggregation_type=args.aggregation_type,
num_scales=args.num_scales,
num_fusions=args.num_fusions,
num_stage_blocks=args.num_stage_blocks,
num_deform_blocks=args.num_deform_blocks,
no_intermediate_supervision=args.no_intermediate_supervision,
refinement_type=args.refinement_type,
mdconv_dilation=args.mdconv_dilation,
deformable_groups=args.deformable_groups).to(device)
# print(aanet)
if os.path.exists(args.pretrained_aanet):
print('=> Loading pretrained AANet:', args.pretrained_aanet)
utils.load_pretrained_net(aanet, args.pretrained_aanet, no_strict=True)
else:
print('=> Using random initialization')
# Save parameters
num_params = utils.count_parameters(aanet)
print('=> Number of trainable parameters: %d' % num_params)
if torch.cuda.device_count() > 1:
print('=> Use %d GPUs' % torch.cuda.device_count())
aanet = torch.nn.DataParallel(aanet)
# Inference
aanet.eval()
inference_time = 0
num_imgs = 0
num_samples = len(test_loader)
print('=> %d samples found in the test set' % num_samples)
for i, sample in enumerate(test_loader):
if args.count_time and i == args.num_images: # testing time only
break
if i % 100 == 0:
print('=> Inferencing %d/%d' % (i, num_samples))
left = sample['left'].to(device) # [B, 3, H, W]
right = sample['right'].to(device)
# Pad
ori_height, ori_width = left.size()[2:]
if ori_height < args.img_height or ori_width < args.img_width:
top_pad = args.img_height - ori_height
right_pad = args.img_width - ori_width
# Pad size: (left_pad, right_pad, top_pad, bottom_pad)
left = F.pad(left, (0, right_pad, top_pad, 0))
right = F.pad(right, (0, right_pad, top_pad, 0))
# Warmup
if i == 0 and args.count_time:
with torch.no_grad():
for _ in range(10):
aanet(left, right)
num_imgs += left.size(0)
with torch.no_grad():
time_start = time.perf_counter()
pred_disp = aanet(left, right)[-1] # [B, H, W]
inference_time += time.perf_counter() - time_start
if pred_disp.size(-1) < left.size(-1):
pred_disp = pred_disp.unsqueeze(1) # [B, 1, H, W]
pred_disp = F.interpolate(pred_disp, (left.size(-2), left.size(-1)),
mode='bilinear', align_corners=True, recompute_scale_factor=True) * (left.size(-1) / pred_disp.size(-1))
pred_disp = pred_disp.squeeze(1) # [B, H, W]
# Crop
if ori_height < args.img_height or ori_width < args.img_width:
if right_pad != 0:
pred_disp = pred_disp[:, top_pad:, :-right_pad]
else:
pred_disp = pred_disp[:, top_pad:]
for b in range(pred_disp.size(0)):
disp = pred_disp[b].detach().cpu().numpy() # [H, W]
save_name = sample['left_name'][b]
save_name = os.path.join(args.output_dir, save_name)
utils.check_path(os.path.dirname(save_name))
if not args.count_time:
if args.save_type == 'pfm':
if args.visualize:
skimage.io.imsave(save_name, (disp * 256.).astype(np.uint16))
save_name = save_name[:-3] + 'pfm'
write_pfm(save_name, disp)
elif args.save_type == 'npy':
save_name = save_name[:-3] + 'npy'
np.save(save_name, disp)
else:
skimage.io.imsave(save_name, (disp * 256.).astype(np.uint16))
print('=> Mean inference time for %d images: %.3fs' % (num_imgs, inference_time / num_imgs))
def main():
# For reproducibility
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
torch.backends.cudnn.benchmark = True
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Test loader
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)
])
aanet = nets.AANet(
args.max_disp,
num_downsample=args.num_downsample,
feature_type=args.feature_type,
no_feature_mdconv=args.no_feature_mdconv,
feature_pyramid=args.feature_pyramid,
feature_pyramid_network=args.feature_pyramid_network,
feature_similarity=args.feature_similarity,
aggregation_type=args.aggregation_type,
num_scales=args.num_scales,
num_fusions=args.num_fusions,
num_stage_blocks=args.num_stage_blocks,
num_deform_blocks=args.num_deform_blocks,
no_intermediate_supervision=args.no_intermediate_supervision,
refinement_type=args.refinement_type,
mdconv_dilation=args.mdconv_dilation,
deformable_groups=args.deformable_groups).to(device)
if os.path.exists(args.pretrained_aanet):
print('=> Loading pretrained AANet:', args.pretrained_aanet)
utils.load_pretrained_net(aanet, args.pretrained_aanet, no_strict=True)
else:
print('=> Using random initialization')
if torch.cuda.device_count() > 1:
print('=> Use %d GPUs' % torch.cuda.device_count())
aanet = torch.nn.DataParallel(aanet)
# Inference
aanet.eval()
if args.data_dir.endswith('/'):
args.data_dir = args.data_dir[:-1]
# all_samples = sorted(glob(args.data_dir + '/*left.png'))
all_samples = sorted(glob(args.data_dir + '/left/*.png'))
num_samples = len(all_samples)
print('=> %d samples found in the data dir' % num_samples)
for i, sample_name in enumerate(all_samples):
if i % 100 == 0:
print('=> Inferencing %d/%d' % (i, num_samples))
left_name = sample_name
right_name = left_name.replace('left', 'right')
left = read_img(left_name)
right = read_img(right_name)
sample = {'left': left, 'right': right}
sample = test_transform(sample) # to tensor and normalize
left = sample['left'].to(device) # [3, H, W]
left = left.unsqueeze(0) # [1, 3, H, W]
right = sample['right'].to(device)
right = right.unsqueeze(0)
# Pad
ori_height, ori_width = left.size()[2:]
# Automatic
factor = 48
args.img_height = math.ceil(ori_height / factor) * factor
args.img_width = math.ceil(ori_width / factor) * factor
if ori_height < args.img_height or ori_width < args.img_width:
top_pad = args.img_height - ori_height
right_pad = args.img_width - ori_width
# Pad size: (left_pad, right_pad, top_pad, bottom_pad)
left = F.pad(left, (0, right_pad, top_pad, 0))
right = F.pad(right, (0, right_pad, top_pad, 0))
with torch.no_grad():
pred_disp = aanet(left, right)[-1] # [B, H, W]
if pred_disp.size(-1) < left.size(-1):
pred_disp = pred_disp.unsqueeze(1) # [B, 1, H, W]
pred_disp = F.interpolate(
pred_disp, (left.size(-2), left.size(-1)),
mode='bilinear') * (left.size(-1) / pred_disp.size(-1))
pred_disp = pred_disp.squeeze(1) # [B, H, W]
# Crop
if ori_height < args.img_height or ori_width < args.img_width:
if right_pad != 0:
pred_disp = pred_disp[:, top_pad:, :-right_pad]
else:
pred_disp = pred_disp[:, top_pad:]
disp = pred_disp[0].detach().cpu().numpy() # [H, W]
save_name = os.path.basename(
left_name)[:-4] + '_' + args.save_suffix + '.png'
save_name = os.path.join(args.output_dir, save_name)
if args.save_type == 'pfm':
if args.visualize:
skimage.io.imsave(save_name, (disp * 256.).astype(np.uint16))
save_name = save_name[:-3] + 'pfm'
write_pfm(save_name, disp)
elif args.save_type == 'npy':
save_name = save_name[:-3] + 'npy'
np.save(save_name, disp)
else:
skimage.io.imsave(save_name, (disp * 256.).astype(np.uint16))
def main():
# For reproducibility
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
torch.backends.cudnn.benchmark = True
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Train loader
train_transform_list = [transforms.RandomCrop(args.img_height, args.img_width),
transforms.RandomColor(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)
]
train_transform = transforms.Compose(train_transform_list)
train_data = dataloader.StereoDataset(data_dir=args.data_dir,
dataset_name=args.dataset_name,
mode='train' if args.mode != 'train_all' else 'train_all',
load_pseudo_gt=args.load_pseudo_gt,
transform=train_transform)
logger.info('=> {} training samples found in the training set'.format(len(train_data)))
train_loader = DataLoader(dataset=train_data, batch_size=args.batch_size, shuffle=True,
num_workers=args.num_workers, pin_memory=True, drop_last=True)
# Validation loader
val_transform_list = [transforms.RandomCrop(args.val_img_height, args.val_img_width, validate=True),
transforms.ToTensor(),
transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)
]
val_transform = transforms.Compose(val_transform_list)
val_data = dataloader.StereoDataset(data_dir=args.data_dir,
dataset_name=args.dataset_name,
mode=args.mode,
transform=val_transform)
val_loader = DataLoader(dataset=val_data, batch_size=args.val_batch_size, shuffle=False,
num_workers=args.num_workers, pin_memory=True, drop_last=False)
# Network
aanet = nets.AANet(args.max_disp,
num_downsample=args.num_downsample,
feature_type=args.feature_type,
no_feature_mdconv=args.no_feature_mdconv,
feature_pyramid=args.feature_pyramid,
feature_pyramid_network=args.feature_pyramid_network,
feature_similarity=args.feature_similarity,
aggregation_type=args.aggregation_type,
num_scales=args.num_scales,
num_fusions=args.num_fusions,
num_stage_blocks=args.num_stage_blocks,
num_deform_blocks=args.num_deform_blocks,
no_intermediate_supervision=args.no_intermediate_supervision,
refinement_type=args.refinement_type,
mdconv_dilation=args.mdconv_dilation,
deformable_groups=args.deformable_groups).to(device)
logger.info('%s' % aanet)
if args.pretrained_aanet is not None:
logger.info('=> Loading pretrained AANet: %s' % args.pretrained_aanet)
# Enable training from a partially pretrained model
utils.load_pretrained_net(aanet, args.pretrained_aanet, no_strict=(not args.strict))
if torch.cuda.device_count() > 1:
logger.info('=> Use %d GPUs' % torch.cuda.device_count())
aanet = torch.nn.DataParallel(aanet)
# Save parameters
num_params = utils.count_parameters(aanet)
logger.info('=> Number of trainable parameters: %d' % num_params)
save_name = '%d_parameters' % num_params
open(os.path.join(args.checkpoint_dir, save_name), 'a').close()
# Optimizer
# Learning rate for offset learning is set 0.1 times those of existing layers
specific_params = list(filter(utils.filter_specific_params,
aanet.named_parameters()))
base_params = list(filter(utils.filter_base_params,
aanet.named_parameters()))
specific_params = [kv[1] for kv in specific_params] # kv is a tuple (key, value)
base_params = [kv[1] for kv in base_params]
specific_lr = args.learning_rate * 0.1
params_group = [
{'params': base_params, 'lr': args.learning_rate},
{'params': specific_params, 'lr': specific_lr},
]
optimizer = torch.optim.Adam(params_group, weight_decay=args.weight_decay)
# Resume training
if args.resume:
# AANet
start_epoch, start_iter, best_epe, best_epoch = utils.resume_latest_ckpt(
args.checkpoint_dir, aanet, 'aanet')
# Optimizer
utils.resume_latest_ckpt(args.checkpoint_dir, optimizer, 'optimizer')
else:
start_epoch = 0
start_iter = 0
best_epe = None
best_epoch = None
# LR scheduler
if args.lr_scheduler_type is not None:
last_epoch = start_epoch if args.resume else start_epoch - 1
if args.lr_scheduler_type == 'MultiStepLR':
milestones = [int(step) for step in args.milestones.split(',')]
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=milestones,
gamma=args.lr_decay_gamma,
last_epoch=last_epoch)
else:
raise NotImplementedError
train_model = model.Model(args, logger, optimizer, aanet, device, start_iter, start_epoch,
best_epe=best_epe, best_epoch=best_epoch)
logger.info('=> Start training...')
if args.evaluate_only:
assert args.val_batch_size == 1
train_model.validate(val_loader)
else:
for _ in range(start_epoch, args.max_epoch):
if not args.evaluate_only:
train_model.train(train_loader)
if not args.no_validate:
train_model.validate(val_loader)
if args.lr_scheduler_type is not None:
lr_scheduler.step()
logger.info('=> End training\n\n')
def main():
# For reproducibility
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
train_loader, val_loader = getDataLoader(args, logger)
# Network
aanet = nets.AANet(
args.max_disp,
num_downsample=args.num_downsample,
feature_type=args.feature_type,
no_feature_mdconv=args.no_feature_mdconv,
feature_pyramid=args.feature_pyramid,
feature_pyramid_network=args.feature_pyramid_network,
feature_similarity=args.feature_similarity,
aggregation_type=args.aggregation_type,
useFeatureAtt=args.useFeatureAtt,
num_scales=args.num_scales,
num_fusions=args.num_fusions,
num_stage_blocks=args.num_stage_blocks,
num_deform_blocks=args.num_deform_blocks,
no_intermediate_supervision=args.no_intermediate_supervision,
refinement_type=args.refinement_type,
mdconv_dilation=args.mdconv_dilation,
deformable_groups=args.deformable_groups).to(device)
# logger.info('%s' % aanet) if local_master else None
if local_master:
structure_of_net = os.path.join(args.checkpoint_dir,
'structure_of_net.txt')
with open(structure_of_net, 'w') as f:
f.write('%s' % aanet)
if args.pretrained_aanet is not None:
logger.info('=> Loading pretrained AANet: %s' % args.pretrained_aanet)
# Enable training from a partially pretrained model
utils.load_pretrained_net(aanet,
args.pretrained_aanet,
no_strict=(not args.strict))
aanet.to(device)
logger.info('=> Use %d GPUs' %
torch.cuda.device_count()) if local_master else None
# if torch.cuda.device_count() > 1:
if args.distributed:
# aanet = torch.nn.DataParallel(aanet)
# 尝试分布式训练
aanet = torch.nn.SyncBatchNorm.convert_sync_batchnorm(aanet)
aanet = torch.nn.parallel.DistributedDataParallel(
aanet, device_ids=[local_rank], output_device=local_rank)
synchronize()
# Save parameters
num_params = utils.count_parameters(aanet)
logger.info('=> Number of trainable parameters: %d' % num_params)
save_name = '%d_parameters' % num_params
open(os.path.join(args.checkpoint_dir, save_name), 'a').close(
) if local_master else None # 这是个空文件，只是通过其文件名称指示模型有多少个需要训练的参数
# Optimizer
# Learning rate for offset learning is set 0.1 times those of existing layers
specific_params = list(
filter(utils.filter_specific_params, aanet.named_parameters()))
base_params = list(
filter(utils.filter_base_params, aanet.named_parameters()))
specific_params = [kv[1]
for kv in specific_params] # kv is a tuple (key, value)
base_params = [kv[1] for kv in base_params]
specific_lr = args.learning_rate * 0.1
params_group = [
{
'params': base_params,
'lr': args.learning_rate
},
{
'params': specific_params,
'lr': specific_lr
},
]
optimizer = torch.optim.Adam(params_group, weight_decay=args.weight_decay)
# Resume training
if args.resume:
# 1. resume AANet
start_epoch, start_iter, best_epe, best_epoch = utils.resume_latest_ckpt(
args.checkpoint_dir, aanet, 'aanet')
# 2. resume Optimizer
utils.resume_latest_ckpt(args.checkpoint_dir, optimizer, 'optimizer')
else:
start_epoch = 0
start_iter = 0
best_epe = None
best_epoch = None
# LR scheduler
if args.lr_scheduler_type is not None:
last_epoch = start_epoch if args.resume else start_epoch - 1
if args.lr_scheduler_type == 'MultiStepLR':
milestones = [int(step) for step in args.milestones.split(',')]
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=milestones,
gamma=args.lr_decay_gamma,
last_epoch=last_epoch
) # 最后这个last_epoch参数很重要：如果是resume的话，则会自动调整学习率适去应last_epoch。
else:
raise NotImplementedError
# model.Model(object)对AANet做了进一步封装。
train_model = model.Model(args,
logger,
optimizer,
aanet,
device,
start_iter,
start_epoch,
best_epe=best_epe,
best_epoch=best_epoch)
logger.info('=> Start training...')
trainLoss_dict, trainLossKey, valLoss_dict, valLossKey = getLossRecord(
netName="AANet")
if args.evaluate_only:
assert args.val_batch_size == 1
train_model.validate(
val_loader, local_master, valLoss_dict,
valLossKey) # test模式。应该设置--evaluate_only，且--mode为“test”。
# 保存Loss用于分析
save_loss_for_matlab(trainLoss_dict, valLoss_dict)
else:
for epoch in range(start_epoch, args.max_epoch): # 训练主循环（Epochs）！！！
if not args.evaluate_only:
# ensure distribute worker sample different data,
# set different random seed by passing epoch to sampler
if args.distributed:
train_loader.sampler.set_epoch(epoch)
logger.info(
'train_loader.sampler.set_epoch({})'.format(epoch))
train_model.train(train_loader, local_master, trainLoss_dict,
trainLossKey)
if not args.no_validate:
train_model.validate(val_loader, local_master, valLoss_dict,
valLossKey) # 训练模式下：边训练边验证。
if args.lr_scheduler_type is not None:
lr_scheduler.step() # 调整Learning Rate
# 保存Loss用于分析。每个epoch结束后，都保存一次，覆盖之前的保存。避免必须训练完成才保存的弊端。
save_loss_for_matlab(trainLoss_dict, valLoss_dict)
logger.info('=> End training\n\n')