def test_from_folders(self, tmpdir):
tmp_dir = Path(tmpdir)
# create random dummy data
os.makedirs(str(tmp_dir / "images"))
os.makedirs(str(tmp_dir / "targets"))
images = [
str(tmp_dir / "images" / "img1.png"),
str(tmp_dir / "images" / "img2.png"),
str(tmp_dir / "images" / "img3.png"),
]
targets = [
str(tmp_dir / "targets" / "img1.png"),
str(tmp_dir / "targets" / "img2.png"),
str(tmp_dir / "targets" / "img3.png"),
]
num_classes: int = 2
img_size: Tuple[int, int] = (196, 196)
create_random_data(images, targets, img_size, num_classes)
# instantiate the data module
dm = SemanticSegmentationData.from_folders(
train_folder=str(tmp_dir / "images"),
train_target_folder=str(tmp_dir / "targets"),
val_folder=str(tmp_dir / "images"),
val_target_folder=str(tmp_dir / "targets"),
test_folder=str(tmp_dir / "images"),
test_target_folder=str(tmp_dir / "targets"),
batch_size=2,
num_workers=0,
num_classes=num_classes,
)
assert dm is not None
assert dm.train_dataloader() is not None
assert dm.val_dataloader() is not None
assert dm.test_dataloader() is not None
# check training data
data = next(iter(dm.train_dataloader()))
imgs, labels = data[DefaultDataKeys.INPUT], data[
DefaultDataKeys.TARGET]
assert imgs.shape == (2, 3, 196, 196)
assert labels.shape == (2, 196, 196)
# check val data
data = next(iter(dm.val_dataloader()))
imgs, labels = data[DefaultDataKeys.INPUT], data[
DefaultDataKeys.TARGET]
assert imgs.shape == (2, 3, 196, 196)
assert labels.shape == (2, 196, 196)
# check test data
data = next(iter(dm.test_dataloader()))
imgs, labels = data[DefaultDataKeys.INPUT], data[
DefaultDataKeys.TARGET]
assert imgs.shape == (2, 3, 196, 196)
assert labels.shape == (2, 196, 196)
import flash
from flash.core.data.utils import download_data
from flash.image import SemanticSegmentation, SemanticSegmentationData
# 1. Create the DataModule
# The data was generated with the CARLA self-driving simulator as part of the Kaggle Lyft Udacity Challenge.
# More info here: https://www.kaggle.com/kumaresanmanickavelu/lyft-udacity-challenge
download_data(
"https://github.com/ongchinkiat/LyftPerceptionChallenge/releases/download/v0.1/carla-capture-20180513A.zip",
"./data",
)
datamodule = SemanticSegmentationData.from_folders(
train_folder="data/CameraRGB",
train_target_folder="data/CameraSeg",
val_split=0.1,
transform_kwargs=dict(image_size=(256, 256)),
num_classes=21,
batch_size=4,
)
# 2. Build the task
model = SemanticSegmentation(
backbone="mobilenetv3_large_100",
head="fpn",
num_classes=datamodule.num_classes,
)
# 3. Create the trainer and finetune the model
trainer = flash.Trainer(max_epochs=3, gpus=torch.cuda.device_count())
trainer.finetune(model, datamodule=datamodule, strategy="freeze")
def test_smoke(self):
dm = SemanticSegmentationData()
assert dm is not None
def test_map_labels(self, tmpdir):
tmp_dir = Path(tmpdir)
# create random dummy data
images = [
str(tmp_dir / "img1.png"),
str(tmp_dir / "img2.png"),
str(tmp_dir / "img3.png"),
]
targets = [
str(tmp_dir / "labels_img1.png"),
str(tmp_dir / "labels_img2.png"),
str(tmp_dir / "labels_img3.png"),
]
labels_map: Dict[int, Tuple[int, int, int]] = {
0: [0, 0, 0],
1: [255, 255, 255],
}
num_classes: int = len(labels_map.keys())
img_size: Tuple[int, int] = (196, 196)
create_random_data(images, targets, img_size, num_classes)
# instantiate the data module
dm = SemanticSegmentationData.from_files(train_files=images,
train_targets=targets,
val_files=images,
val_targets=targets,
batch_size=2,
num_workers=0,
num_classes=num_classes)
assert dm is not None
assert dm.train_dataloader() is not None
# disable visualisation for testing
assert dm.data_fetcher.block_viz_window is True
dm.set_block_viz_window(False)
assert dm.data_fetcher.block_viz_window is False
dm.show_train_batch("load_sample")
dm.show_train_batch("to_tensor_transform")
# check training data
data = next(iter(dm.train_dataloader()))
imgs, labels = data[DefaultDataKeys.INPUT], data[
DefaultDataKeys.TARGET]
assert imgs.shape == (2, 3, 196, 196)
assert labels.shape == (2, 196, 196)
assert labels.min().item() == 0
assert labels.max().item() == 1
assert labels.dtype == torch.int64
# now train with `fast_dev_run`
model = SemanticSegmentation(num_classes=2,
backbone="torchvision/fcn_resnet50")
trainer = Trainer(default_root_dir=tmpdir, fast_dev_run=True)
trainer.finetune(model, dm, strategy="freeze_unfreeze")
import flash
from flash.core.data.utils import download_data
from flash.image import SemanticSegmentation, SemanticSegmentationData
# 1. Create the DataModule
# The data was generated with the CARLA self-driving simulator as part of the Kaggle Lyft Udacity Challenge.
# More info here: https://www.kaggle.com/kumaresanmanickavelu/lyft-udacity-challenge
download_data(
"https://github.com/ongchinkiat/LyftPerceptionChallenge/releases/download/v0.1/carla-capture-20180513A.zip",
"./data",
)
datamodule = SemanticSegmentationData.from_folders(
train_folder="data/CameraRGB",
train_target_folder="data/CameraSeg",
val_split=0.1,
image_size=(256, 256),
num_classes=21,
)
# 2. Build the task
model = SemanticSegmentation(
backbone="mobilenetv3_large_100",
head="fpn",
num_classes=datamodule.num_classes,
)
# 3. Create the trainer and finetune the model
trainer = flash.Trainer(max_epochs=3, gpus=torch.cuda.device_count())
trainer.finetune(model, datamodule=datamodule, strategy="freeze")
def test_smoke():
dm = SemanticSegmentationData(batch_size=1)
assert dm is not None
def test_from_fiftyone(tmpdir):
tmp_dir = Path(tmpdir)
# create random dummy data
images = [
str(tmp_dir / "img1.png"),
str(tmp_dir / "img2.png"),
str(tmp_dir / "img3.png"),
]
num_classes: int = 2
img_size: Tuple[int, int] = (128, 128)
for img_file in images:
_rand_image(img_size).save(img_file)
targets = [
np.array(_rand_labels(img_size, num_classes)) for _ in range(3)
]
dataset = fo.Dataset.from_dir(
str(tmp_dir),
dataset_type=fo.types.ImageDirectory,
)
for idx, sample in enumerate(dataset):
sample["ground_truth"] = fo.Segmentation(mask=targets[idx][:, :,
0])
sample.save()
# instantiate the data module
dm = SemanticSegmentationData.from_fiftyone(
train_dataset=dataset,
val_dataset=dataset,
test_dataset=dataset,
predict_dataset=dataset,
batch_size=2,
num_workers=0,
num_classes=num_classes,
)
assert dm is not None
assert dm.train_dataloader() is not None
assert dm.val_dataloader() is not None
assert dm.test_dataloader() is not None
# check training data
data = next(iter(dm.train_dataloader()))
imgs, labels = data[DataKeys.INPUT], data[DataKeys.TARGET]
assert imgs.shape == (2, 3, 128, 128)
assert labels.shape == (2, 128, 128)
# check val data
data = next(iter(dm.val_dataloader()))
imgs, labels = data[DataKeys.INPUT], data[DataKeys.TARGET]
assert imgs.shape == (2, 3, 128, 128)
assert labels.shape == (2, 128, 128)
# check test data
data = next(iter(dm.test_dataloader()))
imgs, labels = data[DataKeys.INPUT], data[DataKeys.TARGET]
assert imgs.shape == (2, 3, 128, 128)
assert labels.shape == (2, 128, 128)
# check predict data
data = next(iter(dm.predict_dataloader()))
imgs = data[DataKeys.INPUT]
assert imgs.shape == (2, 3, 128, 128)
