def get_dataset_loaders(model, dataset, workers):
target_size = (model["common"]["image_size"],) * 2
batch_size = model["common"]["batch_size"]
path = dataset["common"]["dataset"]
mean, std = [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]
transform = JointCompose(
[
JointTransform(ConvertImageMode("RGB"), ConvertImageMode("P")),
JointTransform(Resize(target_size, Image.BILINEAR), Resize(target_size, Image.NEAREST)),
JointTransform(CenterCrop(target_size), CenterCrop(target_size)),
JointRandomHorizontalFlip(0.5),
JointRandomRotation(0.5, 90),
JointRandomRotation(0.5, 90),
JointRandomRotation(0.5, 90),
JointTransform(ImageToTensor(), MaskToTensor()),
JointTransform(Normalize(mean=mean, std=std), None),
]
)
train_dataset = SlippyMapTilesConcatenation(
[os.path.join(path, "training", "images")], os.path.join(path, "training", "labels"), transform
)
val_dataset = SlippyMapTilesConcatenation(
[os.path.join(path, "validation", "images")], os.path.join(path, "validation", "labels"), transform
)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, drop_last=True, num_workers=workers)
val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, drop_last=True, num_workers=workers)
return train_loader, val_loader
def segment(self, image):
# don't track tensors with autograd during prediction
with torch.no_grad():
mean, std = self.dataset['stats']['mean'], self.dataset['stats'][
'std']
transform = Compose([
ConvertImageMode(mode='RGB'),
ImageToTensor(),
Normalize(mean=mean, std=std)
])
image = transform(image)
batch = image.unsqueeze(0).to(self.device)
output = self.net(batch)
output = output.cpu().data.numpy()
output = output.squeeze(0)
mask = output.argmax(axis=0).astype(np.uint8)
mask = Image.fromarray(mask, mode='P')
palette = make_palette(*self.dataset['common']['colors'])
mask.putpalette(palette)
return mask
def main(args):
dataset = load_config(args.dataset)
path = dataset["common"]["dataset"]
num_classes = len(dataset["common"]["classes"])
train_transform = Compose([ConvertImageMode(mode="P"), MaskToTensor()])
train_dataset = SlippyMapTiles(os.path.join(path, "training", "labels"), transform=train_transform)
n = 0
counts = np.zeros(num_classes, dtype=np.int64)
loader = DataLoader(train_dataset, batch_size=1)
for images, tile in tqdm(loader, desc="Loading", unit="image", ascii=True):
image = torch.squeeze(images)
image = np.array(image, dtype=np.uint8)
n += image.shape[0] * image.shape[1]
counts += np.bincount(image.ravel(), minlength=num_classes)
# Class weighting scheme `w = 1 / ln(c + p)` see:
# - https://arxiv.org/abs/1707.03718
# LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation
# - https://arxiv.org/abs/1606.02147
# ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation
probs = counts / n
weights = 1 / np.log(1.02 + probs)
weights.round(6, out=weights)
print(weights.tolist())
def get_dataset_loaders(model, dataset):
target_size = (model["common"]["image_size"], ) * 2
batch_size = model["common"]["batch_size"]
path = dataset["common"]["dataset"]
mean, std = dataset["stats"]["mean"], dataset["stats"]["std"]
image_transform = Compose([
ConvertImageMode("RGB"),
Resize(target_size, Image.BILINEAR),
CenterCrop(target_size),
ImageToTensor(),
Normalize(mean=mean, std=std),
])
target_transform = Compose([
ConvertImageMode("P"),
Resize(target_size, Image.NEAREST),
CenterCrop(target_size),
MaskToTensor()
])
train_dataset = SlippyMapTilesConcatenation(
[os.path.join(path, "training", "images")],
[image_transform],
os.path.join(path, "training", "labels"),
target_transform,
)
val_dataset = SlippyMapTilesConcatenation(
[os.path.join(path, "validation", "images")],
[image_transform],
os.path.join(path, "validation", "labels"),
target_transform,
)
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
drop_last=True)
val_loader = DataLoader(val_dataset, batch_size=batch_size, drop_last=True)
return train_loader, val_loader
def get_dataset_loaders(model, dataset):
target_size = (model['common']['image_size'], ) * 2
batch_size = model['common']['batch_size']
path = dataset['common']['dataset']
mean, std = dataset['stats']['mean'], dataset['stats']['std']
image_transform = Compose([
ConvertImageMode('RGB'),
Resize(target_size, Image.BILINEAR),
CenterCrop(target_size),
ImageToTensor(),
Normalize(mean=mean, std=std)])
target_transform = Compose([
ConvertImageMode('P'),
Resize(target_size, Image.NEAREST),
CenterCrop(target_size),
MaskToTensor()])
train_dataset = SlippyMapTilesConcatenation(
[os.path.join(path, 'training', 'images')],
[image_transform],
os.path.join(path, 'training', 'labels'),
target_transform)
val_dataset = SlippyMapTilesConcatenation(
[os.path.join(path, 'validation', 'images')],
[image_transform],
os.path.join(path, 'validation', 'labels'),
target_transform)
train_sampler = RandomSubsetSampler(train_dataset, dataset['samples']['training'])
val_sampler = RandomSubsetSampler(val_dataset, dataset['samples']['validation'])
train_loader = DataLoader(train_dataset, batch_size=batch_size, sampler=train_sampler, drop_last=True)
val_loader = DataLoader(val_dataset, batch_size=batch_size, sampler=val_sampler, drop_last=True)
return train_loader, val_loader
def main(args):
dataset = load_config(args.dataset)
path = dataset["common"]["dataset"]
train_transform = Compose([ConvertImageMode(mode="RGB"), ImageToTensor()])
train_dataset = SlippyMapTiles(os.path.join(path, "training", "images"),
transform=train_transform)
n = 0
mean = np.zeros(3, dtype=np.float64)
loader = DataLoader(train_dataset, batch_size=1)
for images, tile in tqdm(loader, desc="Loading", unit="image", ascii=True):
image = torch.squeeze(images)
assert image.size(0) == 3, "channel first"
image = np.array(image, dtype=np.float64)
n += image.shape[1] * image.shape[2]
mean += np.sum(image, axis=(1, 2))
mean /= n
mean.round(decimals=6, out=mean)
print("mean: {}".format(mean.tolist()))
std = np.zeros(3, dtype=np.float64)
loader = DataLoader(train_dataset, batch_size=1)
for images, tile in tqdm(loader, desc="Loading", unit="image", ascii=True):
image = torch.squeeze(images)
assert image.size(0) == 3, "channel first"
image = np.array(image, dtype=np.float64)
difference = np.transpose(image, (1, 2, 0)) - mean
std += np.sum(np.square(difference), axis=(0, 1))
std = np.sqrt(std / (n - 1))
std.round(decimals=6, out=std)
print("std: {}".format(std.tolist()))
def main(args):
model = load_config(args.model)
dataset = load_config(args.dataset)
cuda = model["common"]["cuda"]
device = torch.device("cuda" if cuda else "cpu")
def map_location(storage, _):
return storage.cuda() if cuda else storage.cpu()
if cuda and not torch.cuda.is_available():
sys.exit("Error: CUDA requested but not available")
num_classes = len(dataset["common"]["classes"])
# https://github.com/pytorch/pytorch/issues/7178
chkpt = torch.load(args.checkpoint, map_location=map_location)
net = UNet(num_classes).to(device)
net = nn.DataParallel(net)
if cuda:
torch.backends.cudnn.benchmark = True
net.load_state_dict(chkpt["state_dict"])
net.eval()
mean, std = [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]
transform = Compose([
ConvertImageMode(mode="RGB"),
ImageToTensor(),
Normalize(mean=mean, std=std)
])
directory = BufferedSlippyMapDirectory(args.tiles,
transform=transform,
size=args.tile_size,
overlap=args.overlap)
loader = DataLoader(directory,
batch_size=args.batch_size,
num_workers=args.workers)
# don't track tensors with autograd during prediction
with torch.no_grad():
for images, tiles in tqdm(loader,
desc="Eval",
unit="batch",
ascii=True):
images = images.to(device)
outputs = net(images)
# manually compute segmentation mask class probabilities per pixel
probs = nn.functional.softmax(outputs, dim=1).data.cpu().numpy()
for tile, prob in zip(tiles, probs):
x, y, z = list(map(int, tile))
# we predicted on buffered tiles; now get back probs for original image
prob = directory.unbuffer(prob)
# Quantize the floating point probabilities in [0,1] to [0,255] and store
# a single-channel `.png` file with a continuous color palette attached.
assert prob.shape[
0] == 2, "single channel requires binary model"
assert np.allclose(
np.sum(prob, axis=0), 1.
), "single channel requires probabilities to sum up to one"
foreground = prob[1:, :, :]
anchors = np.linspace(0, 1, 256)
quantized = np.digitize(foreground, anchors).astype(np.uint8)
palette = continuous_palette_for_color("pink", 256)
out = Image.fromarray(quantized.squeeze(), mode="P")
out.putpalette(palette)
os.makedirs(os.path.join(args.probs, str(z), str(x)),
exist_ok=True)
path = os.path.join(args.probs, str(z), str(x),
str(y) + ".png")
out.save(path, optimize=True)
