def transform_tr(self, sample):
composed_transforms = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.RandomScaleCrop(base_size=self.args.base_size,
crop_size=self.args.crop_size,
fill=255),
tr.RandomGaussianBlur(),
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
return composed_transforms(sample)
def transform_tr(self, sample):
composed_transforms = transforms.Compose([
# tr.RandomHorizontalFlip(),
# tr.RandomScaleCrop(base_size=self.args.base_size, crop_size=self.args.crop_size, fill=255),
# tr.RandomCrop2(crop_size=self.args.crop_size),
tr.Rescale(ratio=self.args.ratio),
tr.RandomGaussianBlur(),
tr.Normalize(mean=(0.279, 0.293, 0.290),
std=(0.197, 0.198, 0.201)),
tr.ToTensor()
])
return composed_transforms(sample)
def transform_tr(self, sample):
composed_transforms = transforms.Compose([
# transforms.ColorJitter(brightness=(-1,1),contrast=(-1, 1),saturation=(-0.3, 0.3), hue=(-0.3, 0.3)),
# transforms.ColorJitter(brightness=0.4, contrast=0.4,saturation=0.4),
tr.RandomHorizontalFlip(),
tr.GaussianNoise(),
tr.RandomGaussianBlur(),
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
tr.PatchToTensor()
])
return composed_transforms(sample)
def transform_ts(self, sample):
"""
composed transformers for testing dataset
:param sample: {'image': image, 'label': label}
:return:
"""
composed_transforms = transforms.Compose([
ct.FixScaleCrop(crop_size=self.crop_size),
ct.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
ct.ToTensor()
])
return composed_transforms(sample)
def transform_tr(self, sample):
#print(sample)
composed_transforms = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.FixScaleCrop(crop_size=self.args.crop_size),
#tr.RandomScaleCrop(base_size=self.args.base_size, crop_size=self.args.crop_size),
tr.RandomGaussianBlur(),
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
tr.ToTensor(),
tr.Lablize(high_confidence=self.args.high_confidence)
])
return composed_transforms(sample)
def transform_ts(self, sample):
'''
Transform the given test sample.
@param sample: The given test sample.
'''
composed_transforms = transforms.Compose([
tr.FixedResize(size=self.args.crop_size),
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
tf.ToTensor()
])
return composed_transforms(sample)
def transform_tr(self, sample):
composed_transforms = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.RandomScaleCrop(base_size=513, crop_size=513),
tr.ColorJitter(brightness=0.3,
contrast=0.3,
saturation=0.3,
hue=0.3,
gamma=0.3),
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
return composed_transforms(sample)
def transform_train(self, sample):
composed_transforms = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.RandomVerticalFlip(),
# tr.RandomScaleCrop(base_size=self.args.base_size, crop_size=self.args.crop_size, fill=255),
# tr.FixedResize(size=self.args.crop_size),
tr.RandomRotate(),
tr.RandomGammaTransform(),
tr.RandomGaussianBlur(),
tr.RandomNoise(),
tr.Normalize(mean=(0.544650, 0.352033, 0.384602, 0.352311), std=(0.249456, 0.241652, 0.228824, 0.227583)),
tr.ToTensor()])
return composed_transforms(sample)
def build_data(self):
self.normalize_dict = {
'crop_image': 255.0,
'bifilter_crop_image': 255.0
}
self.composed_transforms_ts = transforms.Compose([
tr.FixedResize({
'image': self.resolution,
'gt': self.resolution
},
flagvals={
'image': cv2.INTER_LINEAR,
'gt': cv2.INTER_NEAREST
}),
tr.CropRandom(crop_size=self.resolution,
keys=['image', 'gt'],
drop_origin=True),
tr.BilateralFiltering(['crop_image']),
tr.ToTensor(),
tr.Normalize('crop_image',
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
])
self.valset = cityscapes_full.cityscapesFullLoader(
split='train_val', transform=self.composed_transforms_ts)
self.valset_num = len(self.valset)
self.val_loader = DataLoader(
self.valset,
batch_size=self.batch_size,
shuffle=False,
num_workers=self.num_workers,
drop_last=True,
)
self.val_loader_num = len(self.val_loader)
self.trainset = cityscapes_full.cityscapesFullLoader(
split='train', transform=self.composed_transforms_ts)
self.trainset_num = len(self.trainset)
self.train_loader = DataLoader(
self.trainset,
batch_size=self.batch_size,
shuffle=True,
num_workers=self.num_workers,
drop_last=True,
)
self.train_loader_num = len(self.train_loader)
def transform_tr(self, sample):
if self.csplit == 'all': ignores = []
elif self.csplit == 'seen': ignores = classes['unseen']
else: raise RuntimeError("Training Unseen data is not legal.")
composed_transforms = transforms.Compose([
tr.MaskIgnores(ignores=ignores, mask=255),
tr.RandomHorizontalFlip(),
# tr.RandomScaleCrop(base_size=self.args.base_size, crop_size=self.args.crop_size),
# tr.RandomGaussianBlur(),
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
seg=True),
tr.ToTensor()
])
return composed_transforms(sample)
def transform_tr(self, sample):
temp = []
if self.args.rotate > 0:
temp.append(tr.RandomRotate(self.args.rotate))
temp.append(tr.RandomScale(rand_resize=self.args.rand_resize))
temp.append(tr.RandomCrop(self.args.input_size))
temp.append(tr.RandomHorizontalFlip())
temp.append(
tr.Normalize(mean=self.args.normal_mean, std=self.args.normal_std))
if self.args.noise_param is not None:
temp.append(
tr.GaussianNoise(mean=self.args.noise_param[0],
std=self.args.noise_param[1]))
temp.append(tr.ToTensor())
composed_transforms = transforms.Compose(temp)
return composed_transforms(sample)
def transform_tr(self, sample):
composed_transforms = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.RandomScaleCrop(base_size=self.args.base_size,
crop_size=self.args.crop_size),
# tr.RandomGaussianBlur(),
# tr.FixedResize(self.args.crop_size),
# tr.RandomCrop(self.args.crop_size),
# tr.RandomCutout(n_holes=1, cut_size=128),
# tr.RandomRotate(30),
tr.RandomRotate_v2(),
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
return composed_transforms(sample)
def explain_image(self, path, counter):
self.model.eval()
img_path = path
lbl_path = os.path.join(
os.path.split(os.path.split(path)[0])[0], 'lbl',
os.path.split(path)[1])
image = Image.open(img_path).convert('RGB') # width x height x 3
_tmp = np.array(Image.open(lbl_path), dtype=np.uint8)
_tmp[_tmp == 255] = 1
_tmp[_tmp == 0] = 0
_tmp[_tmp == 128] = 2
_tmp = Image.fromarray(_tmp)
mean = (0.485, 0.456, 0.406)
std = (0.229, 0.224, 0.225)
composed_transforms = transforms.Compose([
tr.FixedResize(size=513),
tr.Normalize(mean=mean, std=std),
tr.ToTensor()
])
sample = {'image': image, 'label': _tmp}
sample = composed_transforms(sample)
image, target = sample['image'], sample['label']
image = torch.unsqueeze(image, dim=0)
# if self.args.cuda:
# image, target = image.cuda(), target.cuda()
# with torch.no_grad():
# output = self.model(image)
# inn_model = InnvestigateModel(self.model, lrp_exponent=1,
# method="b-rule",
# beta=0, epsilon=1e-6)
#
# inn_model.eval()
# model_prediction, heatmap = inn_model.innvestigate(in_tensor=image)
# model_prediction = np.argmax(model_prediction, axis=1)
# def run_guided_backprop(net, image_tensor):
# return interpretation.guided_backprop(net, image_tensor, cuda=True, verbose=False, apply_softmax=False)
#
# def run_LRP(net, image_tensor):
# return inn_model.innvestigate(in_tensor=image_tensor, rel_for_class=1)
print('hold')
def make_mask(in_flist, model, args):
composed_transforms = transforms.Compose([
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
pbar = tqdm(in_flist, total=len(in_flist))
for fname in pbar:
outpath = os.path.join(args.out_dir, fname)
if os.path.exists(outpath):
pbar.set_description(fname + ' exists')
continue
fpath = os.path.join(args.in_dir, fname)
pbar.set_description('')
image = Image.open(fpath).convert('RGB')
target = image.convert('L')
sample = {'image': image, 'label': target}
tensor_in = composed_transforms(sample)['image'].unsqueeze(0)
if args.cuda:
tensor_in = tensor_in.cuda()
with torch.no_grad():
output = model(tensor_in)
grid_image = make_grid(decode_seg_map_sequence(
torch.max(output[:3], 1)[1].detach().cpu().numpy()),
3,
normalize=False,
range=(0, 255))
grid_image = grid_image[0:1, ...] + grid_image[1:2,
...] + grid_image[2:3,
...]
grid_image[grid_image > 0] = 255
# area = ((grid_image / 255.).sum()) / grid_image.shape[0] / grid_image.shape[1]
_outdir = os.path.split(outpath)[:-1]
_outdir = '/'.join(_outdir)
if not os.path.exists(_outdir):
os.makedirs(_outdir)
assert grid_image.shape[0] == image.size[0]
assert grid_image.shape[1] == image.size[1]
save_image(grid_image, outpath)
def transform_val(self, sample):
if len(sample) == 2:
composed_transforms = transforms.Compose([
tr.FixScaleCrop(crop_size=self.args.crop_size),
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
return composed_transforms(sample)
else:
composed_transforms = transforms.Compose([
tr_sp.FixScaleCrop(crop_size=self.args.crop_size),
tr_sp.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
tr_sp.ToTensor()
])
return composed_transforms(sample)
def transform_tr(self, sample):
color_transforms = [
transforms.RandomApply([transforms.ColorJitter(brightness=0.1)
]), # brightness
transforms.RandomApply([transforms.ColorJitter(contrast=0.1)
]), # contrast
transforms.RandomApply([transforms.ColorJitter(saturation=0.1)
]), # saturation
transforms.RandomApply([transforms.ColorJitter(hue=0.05)])
] # hue
joint_transforms = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.RandomScaleCrop(base_size=self.args.base_size,
crop_size=self.args.crop_size,
fill=255),
tr.equalize(),
tr.RandomGaussianBlur(),
tr.RandomRotate(degree=7)
])
image_transforms = transforms.Compose([
transforms.RandomOrder(color_transforms),
transforms.RandomGrayscale(p=0.3)
])
normalize_transforms = transforms.Compose([
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
tmp_sample = joint_transforms(sample)
tmp_sample['image'] = image_transforms(tmp_sample['image'])
tmp_sample = normalize_transforms(tmp_sample)
return tmp_sample
def transform_val(self, sample):
if self.csplit == 'all':
ignores = []
remap = {}
elif self.csplit == 'unseen':
ignores = classes['seen']
remap = {16: 1, 17: 2, 18: 3, 19: 4, 20: 5}
elif self.csplit == 'seen':
ignores = classes['unseen']
remap = {}
else:
raise RuntimeError("{} ???".format(self.csplit))
composed_transforms = transforms.Compose([
# tr.FixScaleCrop(crop_size=self.args.crop_size),
tr.MaskIgnores(ignores=ignores, mask=0),
tr.ReMask(remap=remap),
tr.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
seg=True),
tr.ToTensor()
])
return composed_transforms(sample)
def transform_tr(self, sample):
"""
composed transformers for training dataset
:param sample: {'image': image, 'label': label}
:return:
"""
img = sample['image']
img = transforms.ColorJitter(brightness=0.5,
contrast=0.5,
saturation=0.5,
hue=0.2)(img)
sample = {'image': img, 'label': sample['label']}
composed_transforms = transforms.Compose([
ct.RandomHorizontalFlip(),
ct.RandomScaleCrop(base_size=self.base_size,
crop_size=self.crop_size),
# ct.RandomChangeBackground(),
ct.RandomGaussianBlur(),
ct.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
ct.ToTensor()
])
return composed_transforms(sample)
device_ids=range(torch.cuda.device_count()))
if resume_epoch != nEpochs:
# Logging into Tensorboard
log_dir = os.path.join(
save_dir, 'models',
datetime.now().strftime('%b%d_%H-%M-%S') + '_' + socket.gethostname())
optimizer = optim.Adam(net.parameters(), lr=p['lr'], weight_decay=p['wd'])
p['optimizer'] = str(optimizer)
composed_transforms_tr = transforms.Compose([
tr.RandomSized(512),
tr.RandomRotate(15),
tr.RandomHorizontalFlip(),
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
composed_transforms_ts = transforms.Compose([
tr.FixedResize(size=(512, 512)),
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
# voc_train = pascal.VOCSegmentation(split='train', transform=composed_transforms_tr)
# voc_val = pascal.VOCSegmentation(split='val', transform=composed_transforms_ts)
ROOT = 'dataset/ORIGA'
voc_train = ImageFolder(root_path=ROOT, datasets='ORIGA')
voc_val = ImageFolder(root_path=ROOT, datasets='ORIGA', mode='test')
def main():
parser = argparse.ArgumentParser(
description="PyTorch DeeplabV3Plus Training")
parser.add_argument('--in-path',
type=str,
required=True,
help='image to test')
# parser.add_argument('--out-path', type=str, required=True, help='mask image to save')
parser.add_argument('--backbone',
type=str,
default='resnet',
choices=['resnet', 'xception', 'drn', 'mobilenet'],
help='backbone name (default: resnet)')
parser.add_argument('--ckpt',
type=str,
default='deeplab-resnet.pth',
help='saved model')
parser.add_argument('--out-stride',
type=int,
default=16,
help='network output stride (default: 8)')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='disables CUDA training')
parser.add_argument('--gpu-ids',
type=str,
default='0',
help='use which gpu to train, must be a \
comma-separated list of integers only (default=0)')
parser.add_argument('--dataset',
type=str,
default='pascal',
choices=['pascal', 'coco', 'cityscapes', 'invoice'],
help='dataset name (default: pascal)')
parser.add_argument('--crop-size',
type=int,
default=513,
help='crop image size')
parser.add_argument('--num_classes',
type=int,
default=2,
help='crop image size')
parser.add_argument('--sync-bn',
type=bool,
default=None,
help='whether to use sync bn (default: auto)')
parser.add_argument(
'--freeze-bn',
type=bool,
default=False,
help='whether to freeze bn parameters (default: False)')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
if args.cuda:
try:
args.gpu_ids = [int(s) for s in args.gpu_ids.split(',')]
except ValueError:
raise ValueError(
'Argument --gpu_ids must be a comma-separated list of integers only'
)
if args.sync_bn is None:
if args.cuda and len(args.gpu_ids) > 1:
args.sync_bn = True
else:
args.sync_bn = False
model_s_time = time.time()
model = DeepLab(num_classes=args.num_classes,
backbone=args.backbone,
output_stride=args.out_stride,
sync_bn=args.sync_bn,
freeze_bn=args.freeze_bn)
ckpt = torch.load(args.ckpt, map_location='cpu')
model.load_state_dict(ckpt['state_dict'])
model = model.cuda()
model_u_time = time.time()
model_load_time = model_u_time - model_s_time
print("model load time is {}".format(model_load_time))
composed_transforms = transforms.Compose([
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
for name in os.listdir(args.in_path):
s_time = time.time()
image = Image.open(args.in_path + "/" + name).convert('RGB')
# image = Image.open(args.in_path).convert('RGB')
target = Image.open(args.in_path + "/" + name).convert('L')
sample = {'image': image, 'label': target}
tensor_in = composed_transforms(sample)['image'].unsqueeze(0)
model.eval()
if args.cuda:
tensor_in = tensor_in.cuda()
with torch.no_grad():
output = model(tensor_in)
grid_image = make_grid(decode_seg_map_sequence(
torch.max(output[:3], 1)[1].detach().cpu().numpy()),
3,
normalize=False,
range=(0, 255))
save_image(grid_image,
args.in_path + "/" + "{}_mask.png".format(name[0:-4]))
u_time = time.time()
img_time = u_time - s_time
print("image:{} time: {} ".format(name, img_time))
# save_image(grid_image, args.out_path)
# print("type(grid) is: ", type(grid_image))
# print("grid_image.shape is: ", grid_image.shape)
print("image save in in_path.")
return _img, _target
def __str__(self):
return 'SBDSegmentation(split=' + str(self.split) + ')'
if __name__ == '__main__':
from dataloaders import custom_transforms as tr
from torch.utils.data import DataLoader
from torchvision import transforms
import matplotlib.pyplot as plt
composed_transforms_tr = transforms.Compose([
tr.RandomResizedCrop(size=513),
tr.RandomHorizontalFlip(),
tr.Normalize(mean=(0.4914, 0.4822, 0.4465), std=(1.0, 1.0, 1.0)),
tr.ToTensor()
])
sbd_train = SBDSegmentation(split='train',
transform=composed_transforms_tr)
dataloader = DataLoader(sbd_train,
batch_size=2,
shuffle=True,
num_workers=2)
for ii, sample in enumerate(dataloader):
for jj in range(sample["image"].size()[0]):
img = sample['image'].numpy()
gt = sample['gt'].numpy()
from torchvision import transforms
transform = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.ScaleNRotate(rots=(-20, 20), scales=(.75, 1.25)),
# tr.CropFromMask(crop_elems=('image', 'gt'), relax=30, zero_pad=False, jitters_bound=(10, 30)),
tr.CropFromMask(crop_elems=('image', 'gt'),
relax=30,
zero_pad=False,
jitters_bound=(30, 31)),
# tr.CropFromMask(crop_elems=('image', 'gt'), relax=30, zero_pad=False, jitters_bound=None),
tr.FixedResize(resolutions={
'crop_image': (256, 256),
'crop_gt': (256, 256)
}),
tr.Normalize(elems='crop_image'),
# tr.ToImage(norm_elem=('pos_map', 'neg_map')),
])
dataset = VOCSegmentation(split=['val'], transform=transform, retname=True)
# dataset = VOCSegmentation(split=['train', 'val'], retname=True)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=0)
for i, sample in enumerate(dataloader):
while 1:
cv2.imshow('image', sample['crop_image'][0].numpy())
cv2.imshow('crop_gt', sample['crop_gt'][0].numpy())
if cv2.waitKey(1) & 0xff == ord('q'):
def main():
parser = argparse.ArgumentParser(
description="PyTorch DeeplabV3Plus Training")
parser.add_argument('--in-path',
type=str,
required=True,
help='directory of images to test')
parser.add_argument('--out-path',
type=str,
required=True,
help='directory of mask image to save')
parser.add_argument('--backbone',
type=str,
default='resnet',
choices=['resnet', 'xception', 'drn', 'mobilenet'],
help='backbone name (default: resnet)')
parser.add_argument('--ckpt',
type=str,
default='deeplab-resnet.pth',
help='saved model')
parser.add_argument('--out-stride',
type=int,
default=16,
help='network output stride (default: 8)')
parser.add_argument('--no-cuda',
action='store_true',
default=True,
help='disables CUDA training')
parser.add_argument('--gpu-ids',
type=str,
default='0',
help='use which gpu to train, must be a \
comma-separated list of integers only (default=0)')
parser.add_argument('--dataset',
type=str,
default='pascal',
choices=['pascal', 'coco', 'cityscapes'],
help='dataset name (default: pascal)')
parser.add_argument('--crop-size',
type=int,
default=512,
help='crop image size')
parser.add_argument('--sync-bn',
type=bool,
default=None,
help='whether to use sync bn (default: auto)')
parser.add_argument(
'--freeze-bn',
type=bool,
default=False,
help='whether to freeze bn parameters (default: False)')
args = parser.parse_args()
print(args)
args.cuda = not args.no_cuda and torch.cuda.is_available()
if args.cuda:
try:
args.gpu_ids = [int(s) for s in args.gpu_ids.split(',')]
except ValueError:
raise ValueError(
'Argument --gpu_ids must be a comma-separated list of integers only'
)
if args.sync_bn is None:
if args.cuda and len(args.gpu_ids) > 1:
args.sync_bn = True
else:
args.sync_bn = False
model = DeepLab(num_classes=3,
backbone=args.backbone,
output_stride=args.out_stride,
sync_bn=args.sync_bn,
freeze_bn=args.freeze_bn)
print(f"The args.ckpt is : {args.ckpt}")
ckpt = torch.load(args.ckpt, map_location='cpu')
model.load_state_dict(ckpt['state_dict'])
composed_transforms = transforms.Compose([
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
for img_path in os.listdir(args.in_path):
if os.path.splitext(img_path)[-1] not in ['.jpg']:
print('skip {}'.format(img_path))
continue
img_path = os.path.join(args.in_path, img_path)
output_path = os.path.join(
args.out_path,
os.path.splitext(os.path.split(img_path)[-1])[-2] + "-seg" +
".jpg")
# print("output path is {}".format(output_path))
combine_path = os.path.join(
args.out_path,
os.path.splitext(os.path.split(img_path)[-1])[-2] + "-blend" +
".png")
# print("blend path is {}".format(combine_path))
image = Image.open(img_path).convert('RGB')
target = Image.open(img_path).convert('L')
sample = {'image': image, 'label': target}
tensor_in = composed_transforms(sample)['image'].unsqueeze(0)
model.eval()
if args.cuda:
image = image.cuda()
with torch.no_grad():
output = model(tensor_in)
grid_image = make_grid(decode_seg_map_sequence(
torch.max(output[:3], 1)[1].detach().cpu().numpy()),
3,
normalize=False,
range=(0, 255))
print("type(grid) is:{}".format(type(grid_image)))
print("grid_image.shape is:{}".format(grid_image.shape))
save_image(grid_image, output_path)
print("saved {}".format(output_path))
blend_two_images(img_path, output_path, combine_path)
print("blended {}\n".format(combine_path))
def transform_norm(self, sample):
composed_transforms = transforms.Compose([
tr.Normalize(mean=self.data_mean, std=self.data_std),
tr.ToTensor()
])
return composed_transforms(sample)
def main():
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
if args.cuda:
try:
args.gpu_ids = [int(s) for s in args.gpu_ids.split(',')]
except ValueError:
raise ValueError(
'Argument --gpu_ids must be a comma-separated list of integers only'
)
if args.sync_bn is None:
if args.cuda and len(args.gpu_ids) > 1:
args.sync_bn = True
else:
args.sync_bn = False
model = DeepLab(num_classes=21,
backbone=args.backbone,
output_stride=args.out_stride,
sync_bn=args.sync_bn,
freeze_bn=args.freeze_bn)
ckpt = torch.load(args.ckpt, map_location='cpu')
model.load_state_dict(ckpt['state_dict'])
composed_transforms = transforms.Compose([
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
# 打开TXT，获取待测试的图片名
filename = args.test_folder + 'test.txt'
file = open(filename)
for line in file:
# 生成输入图片的路径和存储的路径
image_name = line.strip('\n')
input_image = args.test_folder + image_name + '.jpg'
print(input_image)
save_result = args.save_testresult_folder + 'result_' + image_name + '.jpg'
print(save_result)
# 开始测试~~
image = Image.open(input_image).convert('RGB')
target = Image.open(input_image).convert('L')
sample = {'image': image, 'label': target}
tensor_in = composed_transforms(sample)['image'].unsqueeze(0)
model.eval()
if args.cuda:
image = image.cuda()
with torch.no_grad():
output = model(tensor_in)
grid_image = make_grid(decode_seg_map_sequence(
torch.max(output[:3], 1)[1].detach().cpu().numpy()),
3,
normalize=False,
range=(0, 255))
print("type(grid) is: ", type(grid_image))
print("grid_image.shape is: ", grid_image.shape)
save_image(grid_image, save_result)
def main():
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
# args.cuda = torch.cuda.is_available()
if args.cuda:
try:
args.gpu_ids = [int(s) for s in args.gpu_ids.split(',')]
except ValueError:
raise ValueError(
'Argument --gpu_ids must be a comma-separated list of integers only'
)
if args.sync_bn is None:
if args.cuda and len(args.gpu_ids) > 1:
args.sync_bn = True
else:
args.sync_bn = False
model = DeepLab(num_classes=21,
backbone=args.backbone,
output_stride=args.out_stride,
sync_bn=args.sync_bn,
freeze_bn=args.freeze_bn)
ckpt = torch.load(args.ckpt, map_location='cpu')
model.load_state_dict(ckpt['state_dict'])
composed_transforms = transforms.Compose([
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
#
filename = args.test_folder + 'test.txt'
file = open(filename)
# init time parameters
forward_min = math.inf
forward_time = 0
model_time_min = math.inf
model_time = 0
image_number = 0
# init model cuda and image index
model.eval()
image_width = 1920
image_heigth = 1080
if args.cuda:
model.cuda() # 第一次挂载到CUDA需要消耗很长时间，大概4.1208秒
index = torch.arange(1, image_heigth + 1).t().reshape(
-1, 1) # 生成1-1080的数组 代表一行中的每一个像素的索引
# print("index = ", index, index.size())
img_index = index.repeat(
1, image_width).cuda().float() # 重复1920次，得到1080行，1920列的数组
# print("img_index = ", img_index[0], img_index, img_index.size())
else:
index = torch.arange(1, image_heigth +
1).t() # 生成1-1000的数组 代表一行中的每一个像素的索引
img_index = index.repeat(image_width, 1).reshape(
image_heigth, image_width) # 重复838次，得到838行，1000列的数组
for line in file:
# 生成输入图片的路径和存储的路径
image_name = line.strip('\n')
input_image = args.test_folder + image_name + '.jpg'
print(input_image)
save_result = args.save_testresult_folder + 'result_' + image_name + '.jpg'
print(save_result)
image_number = image_number + 1
# 开始测试~~
image = Image.open(input_image).convert('RGB')
target = Image.open(input_image).convert('L')
sample = {'image': image, 'label': target}
tensor_in = composed_transforms(sample)['image'].unsqueeze(0)
# 图片转换成数组，两个数组是因为需要大光斑（红色）和小光斑（绿色）
image_array = np.array(image) # 存储大光斑（红色）
image_array2 = np.copy(image_array) # 存储小光斑（绿色）
image_tensor = torch.from_numpy(image_array)
# image_tensor2 = torch.from_numpy(image_array2)
# model.eval()
if args.cuda:
# model.cuda() # 第一次挂载到CUDA需要消耗很长时间，大概4.1208秒
tensor_in = tensor_in.cuda()
image_tensor = image_tensor.cuda()
# image_tensor2 = image_tensor2.cuda()
start = time.time()
with torch.no_grad():
output = model(tensor_in) #.cpu()
# convert to TensorRT feeding sample data as input
# model_trt = torch2trt(model, [tensor_in])
# output = model_trt(tensor_in).cpu()
# find background index
# torch.max output max value as dim 0, and index of max value as dim 1; use this we can get the class of each pixel
image_tensor[torch.max(output[:3], 1)[1][0] != 1] = 0
# get red channel and use red channel to calculate gray gravity
image_tensor_gray = image_tensor[:, :, 0] # red channel
# save_result_roi = args.save_testresult_folder + 'result_roitemp' + image_name + '.jpg'
# cv2.imwrite(save_result_roi, image_tensor_gray.cpu().numpy())
elapsed = time.time() - start
model_time_min = min(model_time_min, elapsed)
model_time += elapsed
# claculate gray center pixel
# roi_image = cv2.merge([image_tensor.cpu().numpy()])
# gray_center = Cvpointgray(roi_image)
# gray_center = Cvpointgray_fast(roi_image)
gray_center = Cvpointgray_gpu(image_tensor_gray, img_index)
elapsed = time.time() - start
forward_min = min(forward_min, elapsed)
forward_time += elapsed
print("elapsed = ", elapsed)
# 再读一次图片，用作展示
image_orign = cv2.imread(
input_image) #Image.open(input_image).convert('RGB')
gray_center = gray_center.cpu().numpy()
print("len(gray_center) = ", len(gray_center))
for i in range(len(gray_center)):
if np.isnan(gray_center[i]) or np.isinf(gray_center[i]):
continue
else:
cv2.circle(image_orign, (int(i), int(gray_center[i])), 2,
(0, 255, 0))
# 画大光条，易于理解的灰度重心法
# for i in range(len(gray_center)):
# cv2.circle(image_orign, ( int(gray_center[i][0]), int(gray_center[i][1]) ), 2, (0,0,255))
save_result_fin = args.save_testresult_folder + 'result_fin_' + image_name + '.jpg'
cv2.imwrite(save_result_fin, image_orign)
# test time
# lp = LineProfiler()
# lp_wrapper = lp(Cvpointgray_gpu)
# lp_wrapper(image_tensor_gray, img_index)
# lp.print_stats()
# save class heatmap
grid_image = make_grid(decode_seg_map_sequence(
torch.max(output[:3], 1)[1].detach().cpu().numpy()),
3,
normalize=False,
range=(0, 255))
print("type(grid) is: ", type(grid_image))
print("grid_image.shape is: ", grid_image.shape)
save_image(grid_image, save_result)
forward_average = forward_time / image_number
model_average = model_time / image_number
print('Forward: {0:.3f}/{1:.3f}'.format(forward_min, forward_average), 's')
print('Model: {0:.3f}/{1:.3f}'.format(model_time_min, model_average), 's')
def pred_single_image(self, path, counter):
self.model.eval()
img_path = path
lbl_path = os.path.join(
os.path.split(os.path.split(path)[0])[0], 'lbl',
os.path.split(path)[1])
activations = collections.defaultdict(list)
def save_activation(name, mod, input, output):
activations[name].append(output.cpu())
for name, m in self.model.named_modules():
if type(m) == nn.ReLU:
m.register_forward_hook(partial(save_activation, name))
input = cv2.imread(path)
# bkg = cv2.createBackgroundSubtractorMOG2()
# back = bkg.apply(input)
# cv2.imshow('back', back)
# cv2.waitKey()
input = cv2.resize(input, (513, 513), interpolation=cv2.INTER_CUBIC)
image = Image.open(img_path).convert('RGB') # width x height x 3
# _tmp = np.array(Image.open(lbl_path), dtype=np.uint8)
_tmp = np.array(Image.open(img_path), dtype=np.uint8)
_tmp[_tmp == 255] = 1
_tmp[_tmp == 0] = 0
_tmp[_tmp == 128] = 2
_tmp = Image.fromarray(_tmp)
mean = (0.485, 0.456, 0.406)
std = (0.229, 0.224, 0.225)
composed_transforms = transforms.Compose([
tr.FixedResize(size=513),
tr.Normalize(mean=mean, std=std),
tr.ToTensor()
])
sample = {'image': image, 'label': _tmp}
sample = composed_transforms(sample)
image, target = sample['image'], sample['label']
image = torch.unsqueeze(image, dim=0)
if self.args.cuda:
image, target = image.cuda(), target.cuda()
with torch.no_grad():
output = self.model(image)
see = Analysis('module.decoder.last_conv.6', activations)
pred = output.data.cpu().numpy()
target = target.cpu().numpy()
pred = np.argmax(pred, axis=1)
pred = np.reshape(pred, (513, 513))
# prediction = np.append(target, pred, axis=1)
prediction = pred
rgb = np.zeros((prediction.shape[0], prediction.shape[1], 3))
r = prediction.copy()
g = prediction.copy()
b = prediction.copy()
g[g != 1] = 0
g[g == 1] = 255
r[r != 2] = 0
r[r == 2] = 255
b = np.zeros(b.shape)
rgb[:, :, 0] = b
rgb[:, :, 1] = g
rgb[:, :, 2] = r
prediction = np.append(input, rgb.astype(np.uint8), axis=1)
result = np.append(input, prediction.astype(np.uint8), axis=1)
cv2.line(rgb, (513, 0), (513, 1020), (255, 255, 255), thickness=1)
cv2.line(rgb, (513, 0), (513, 1020), (255, 255, 255), thickness=1)
cv2.imwrite(
'/home/robot/git/pytorch-deeplab-xception/run/cropweed/deeplab-resnet/experiment_41/samples/synthetic_{}.png'
.format(counter), prediction)
def pred_single_image(self, path):
self.model.eval()
img_path = path
lbl_path = os.path.join(
os.path.split(os.path.split(path)[0])[0], 'lbl',
os.path.split(path)[1])
activations = collections.defaultdict(list)
def save_activation(name, mod, input, output):
activations[name].append(output.cpu())
for name, m in self.model.named_modules():
if type(m) == nn.ReLU:
m.register_forward_hook(partial(save_activation, name))
input = cv2.imread(path)
label = cv2.imread(lbl_path)
# bkg = cv2.createBackgroundSubtractorMOG2()
# back = bkg.apply(input)
# cv2.imshow('back', back)
# cv2.waitKey()
input = cv2.resize(input, (513, 513), interpolation=cv2.INTER_CUBIC)
image = Image.open(img_path).convert('RGB') # width x height x 3
# _tmp = np.array(Image.open(lbl_path), dtype=np.uint8)
_tmp = np.array(Image.open(img_path), dtype=np.uint8)
_tmp[_tmp == 255] = 1
_tmp[_tmp == 0] = 0
_tmp[_tmp == 128] = 2
_tmp = Image.fromarray(_tmp)
mean = (0.485, 0.456, 0.406)
std = (0.229, 0.224, 0.225)
composed_transforms = transforms.Compose([
tr.FixedResize(size=513),
tr.Normalize(mean=mean, std=std),
tr.ToTensor()
])
sample = {'image': image, 'label': _tmp}
sample = composed_transforms(sample)
image, target = sample['image'], sample['label']
image = torch.unsqueeze(image, dim=0)
if self.args.cuda:
image, target = image.cuda(), target.cuda()
with torch.no_grad():
output = self.model(image)
# output = output.data.cpu().numpy().squeeze(0).transpose([1, 2, 0])
# output = np.argmax(output, axis=2) * 255
output = output.data.cpu().numpy()
prediction = np.argmax(output, axis=1)
prediction = np.squeeze(prediction, axis=0)
prediction[prediction == 1] = 255
if np.any(prediction == 2):
prediction[prediction == 2] = 128
if np.any(prediction == 1):
prediction[prediction == 1] = 255
print(np.unique(prediction))
see = Analysis(activations, label=1, path=self.saver.experiment_dir)
see.backtrace(output)
# for key in keys:
#
# see.visualize_tensor(see.image)
# see.save_tensor(see.image, self.saver.experiment_dir)
cv2.imwrite(os.path.join(self.saver.experiment_dir, 'rgb.png'), input)
cv2.imwrite(os.path.join(self.saver.experiment_dir, 'lbl.png'), label)
cv2.imwrite(os.path.join(self.saver.experiment_dir, 'prediction.png'),
prediction)
def __init__(self,
base_dir='/home/timy90022/dataset/istg/',
split='train'
):
"""
:param split: train/val/test
"""
super().__init__()
self._base_dir = base_dir
self._image_dir = os.path.join(self._base_dir, 'JPEGImages')
self.rotate = True
self.gaussian_bump = True
self.gaussian_rad = -1
self.gaussian_iou = 0.7
self.split = split
self.max_ratio = 1
self.min_ratio = 1
self.record_ratio_door = []
self.record_ratio_window = []
self.record_area_door = []
self.record_area_window = []
self.mean = (0.5,0.5,0.5)
self.std = (0.5,0.5,0.5)
self.transform = composed_transforms_tr = transforms.Compose([
tr.Normalize(mean=(0.5,0.5,0.5), std=(0.5,0.5,0.5))
])
self.output_size = [128,256]
self.input_size = [512,1024]
self.width_ratio = self.output_size[1] / self.input_size[1]
self.height_ratio = self.output_size[0] / self.input_size[0]
self.all_image = []
self.all_label = []
self.all_assist = []
first = sorted(glob.glob(self._base_dir + '/data_obj2d/istg' + '/*'))
for i in first:
second = sorted(glob.glob(i + '/*'))
for j in second:
# print(j)
file_path_one = j + '/color.png'
file_path_two = j + '/obj2d.png'
file_path_three = j.replace('data_obj2d','fcmaps') + '/fcmap.png'
# print(os.path.isfile(file_path_two))
# print(file_path_two)
if ( os.path.isfile(file_path_one) and os.path.isfile(file_path_two) and os.path.isfile(file_path_three)):
self.all_image.append(file_path_one)
self.all_label.append(file_path_two)
self.all_assist.append(file_path_three)
first = sorted(glob.glob(self._base_dir + '/data_obj2d/sun360' + '/*'))
for i in first:
# print(j)
file_path_one = i + '/color.png'
file_path_two = i + '/obj2d.png'
file_path_three = i.replace('data_obj2d','fcmaps') + '/fcmap.png'
# print(os.path.isfile(file_path_two))
# print(file_path_two)
if ( os.path.isfile(file_path_one) and os.path.isfile(file_path_two) and os.path.isfile(file_path_three)):
self.all_image.append(file_path_one)
self.all_label.append(file_path_two)
self.all_assist.append(file_path_three)
assert (len(self.all_image) == len(self.all_label))
if self.split =='train':
stay = [0,1,2,3,4,5,6,7,8]
else:
stay = [9]
self.all_image = [self.all_image[e] for e in range(len(self.all_image)) if e%10 in stay]
self.all_label = [self.all_label[e] for e in range(len(self.all_label)) if e%10 in stay]
self.all_assist = [self.all_assist[e] for e in range(len(self.all_assist)) if e%10 in stay]
# Display stats
print('Number of images in {}: {:d}'.format(split, len(self.all_image)))
def main():
parser = argparse.ArgumentParser(
description="PyTorch DeeplabV3Plus Training")
parser.add_argument('--in-path',
type=str,
required=True,
help='image folder path to test')
parser.add_argument('--out-path',
type=str,
default='inference_results',
help='mask image folder to save')
parser.add_argument('--backbone',
type=str,
default='resnet',
choices=['resnet'],
help='backbone name (default: resnet)')
parser.add_argument('--ckpt',
type=str,
default='modeling/deeplab-resnet.pth.tar',
help='saved model')
parser.add_argument('--out-stride',
type=int,
default=16,
help='network output stride (default: 16)')
parser.add_argument('--no-cuda',
action='store_true',
default=True,
help='disables CUDA training')
parser.add_argument('--gpu-ids',
type=str,
default='0',
help='use which gpu to train, must be a \
comma-separated list of integers only (default=0)')
parser.add_argument('--dataset',
type=str,
default='coco',
choices=['coco'],
help='dataset name (default: coco)')
parser.add_argument('--crop-size',
type=int,
default=513,
help='crop image size')
parser.add_argument('--sync-bn',
type=bool,
default=None,
help='whether to use sync bn (default: auto)')
parser.add_argument(
'--freeze-bn',
type=bool,
default=False,
help='whether to freeze bn parameters (default: False)')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
if args.cuda:
try:
args.gpu_ids = [int(s) for s in args.gpu_ids.split(',')]
except ValueError:
raise ValueError(
'Argument --gpu_ids must be a comma-separated list of integers only'
)
if args.sync_bn is None:
if args.cuda and len(args.gpu_ids) > 1:
args.sync_bn = True
else:
args.sync_bn = False
model = DeepLab(num_classes=21,
backbone=args.backbone,
output_stride=args.out_stride,
sync_bn=args.sync_bn,
freeze_bn=args.freeze_bn)
ckpt = torch.load(args.ckpt, map_location='cpu')
model.load_state_dict(ckpt['state_dict'])
composed_transforms = transforms.Compose([
tr.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
tr.ToTensor()
])
if not os.path.exists(args.out_path):
os.makedirs(args.out_path)
for img_path in glob.glob(args.in_path + '/*.jpg'):
image = Image.open(img_path).convert('RGB')
target = Image.open(img_path).convert('L')
sample = {'image': image, 'label': target}
tensor_in = composed_transforms(sample)['image'].unsqueeze(0)
model.eval()
if args.cuda:
image = image.cuda()
with torch.no_grad():
output = model(tensor_in)
grid_image = make_grid(decode_seg_map_sequence(
torch.max(output[:3], 1)[1].detach().cpu().numpy()),
3,
normalize=False,
range=(0, 255))
print("type(grid) is: ", type(grid_image))
print("grid_image.shape is: ", grid_image.shape)
img_save = img_path.replace(args.in_path, '')
img_save = 'inference_' + img_save[1:]
save_image(grid_image, args.out_path + '/' + img_save)
