Add SimCLR trainer (#1252)

* add simclr and tests

* add lightly to reqs

* pyupgrade

* Copy things from prior implementation

* Add SimCLR v2 projection head

* Remove kwargs

* Call __init__ explicitly

* Fix mypy and docs

* Can't test newer setuptools

* Default to output dim of model

* Add memory bank

* Ignore erroneous warning

* Fix configs, test SSL4EO

* Fix a few layer bugs

* mypy fixes

* kernel_size must be an integer

* Fix SeCo in_channels

* Get more coverage

* Bump min lightly

* Default logging

* Test weights

* mypy fix

* Grab max_epochs from the trainer

* max_epochs param removed

* Use num_features

* Remove classification head

* SimCLR uses LARS, with Adam as a backup

* Add warnings

* Grab num features directly from model

* Check if identity

* Match timm model design

* Capture warnings

* Fix tests

* Increase coverage

* Fix method name

* More typos

* Escape regex

* Newer setuptools now supported

* New batch norm for every layer

* Rename forward arg

* Clarify usage of weights parameter

Co-authored-by: Caleb Robinson <calebrob6@gmail.com>

* Fix flake8

* Check it

* Use hydra

* Track average L2 normed stdev over features

* SimCLR decays lr to 0

* Add lr warmup

* Fix version access

* Fix LinearLR

* isinstance supports tuples

* Comment capitalization

* Require lightly 1.4.3+

* Require lightly 1.4.3+

* Bump lightly version

* Add RandomGrayscale

* Flake8 fixes

* Placate pydocstyle

* Clarify docs

* Pass correct weights

---------

Co-authored-by: Adam J. Stewart <ajstewart426@gmail.com>
Co-authored-by: Caleb Robinson <calebrob6@gmail.com>

docs/conf.py

@@ -62,6 +62,7 @@ nitpick_ignore = [
     ("py:class", "segmentation_models_pytorch.base.model.SegmentationModel"),
     ("py:class", "timm.models.resnet.ResNet"),
     ("py:class", "timm.models.vision_transformer.VisionTransformer"),
+    ("py:class", "torch.optim.lr_scheduler.LRScheduler"),
     ("py:class", "torchvision.models._api.WeightsEnum"),
     ("py:class", "torchvision.models.resnet.ResNet"),
 ]

environment.yml

@@ -26,6 +26,7 @@ dependencies:
     - isort[colors]>=5.8
     - kornia>=0.6.5
     - laspy>=2
+    - lightly>=1.4.4
     - lightning>=1.8
     - mypy>=0.900
     - nbmake>=1.3.3

requirements/min-reqs.old

@@ -5,6 +5,7 @@ setuptools==42.0.0
 einops==0.3.0
 fiona==1.8.19
 kornia==0.6.5
+lightly==1.4.4
 lightning==1.8.0
 matplotlib==3.3.3
 numpy==1.19.3

requirements/required.txt

@@ -6,6 +6,7 @@ einops==0.6.1
 fiona==1.9.3
 kornia==0.6.12
 lightning==2.0.2
+lightly==1.4.4
 matplotlib==3.7.1
 numpy==1.24.3
 pillow==9.5.0

setup.cfg

@@ -30,6 +30,8 @@ install_requires =
     fiona>=1.8.19,<2
     # kornia 0.6.5+ required due to change in kornia.augmentation API
     kornia>=0.6.5,<0.7
+    # lightly 1.4.4+ required for MoCo v3 support
+    lightly>=1.4.4
     # lightning 1.8+ is first release
     lightning>=1.8,<3
     # matplotlib 3.3.3+ required for Python 3.9 wheels

tests/conf/chesapeake_cvpr_prior_simclr.yaml

@@ -0,0 +1,23 @@
+module:
+  _target_: torchgeo.trainers.SimCLRTask
+  model: "resnet18"
+  in_channels: 4
+  version: 1
+  layers: 2
+  memory_bank_size: 0
+
+datamodule:
+  _target_: torchgeo.datamodules.ChesapeakeCVPRDataModule
+  root: "tests/data/chesapeake/cvpr"
+  download: false
+  train_splits:
+  - "de-test"
+  val_splits:
+  - "de-test"
+  test_splits:
+  - "de-test"
+  batch_size: 2
+  patch_size: 64
+  num_workers: 0
+  class_set: 5
+  use_prior_labels: True

tests/conf/seco_simclr_1.yaml

@@ -0,0 +1,17 @@
+module:
+  _target_: torchgeo.trainers.SimCLRTask
+  model: "resnet18"
+  in_channels: 3
+  version: 1
+  layers: 2
+  hidden_dim: 8
+  output_dim: 8
+  weight_decay: 1e-6
+  memory_bank_size: 0
+
+datamodule:
+  _target_: torchgeo.datamodules.SeasonalContrastS2DataModule
+  root: "tests/data/seco"
+  seasons: 1
+  batch_size: 2
+  num_workers: 0

tests/conf/seco_simclr_2.yaml

@@ -0,0 +1,17 @@
+module:
+  _target_: torchgeo.trainers.SimCLRTask
+  model: "resnet18"
+  in_channels: 3
+  version: 2
+  layers: 4
+  hidden_dim: 8
+  output_dim: 8
+  weight_decay: 1e-4
+  memory_bank_size: 10
+
+datamodule:
+  _target_: torchgeo.datamodules.SeasonalContrastS2DataModule
+  root: "tests/data/seco"
+  seasons: 2
+  batch_size: 2
+  num_workers: 0

tests/conf/ssl4eo_s12_simclr_1.yaml

@@ -0,0 +1,17 @@
+module:
+  _target_: torchgeo.trainers.SimCLRTask
+  model: "resnet18"
+  in_channels: 13
+  version: 1
+  layers: 2
+  hidden_dim: 8
+  output_dim: 8
+  weight_decay: 1e-6
+  memory_bank_size: 0
+
+datamodule:
+  _target_: torchgeo.datamodules.SSL4EOS12DataModule
+  root: "tests/data/ssl4eo/s12"
+  seasons: 1
+  batch_size: 2
+  num_workers: 0

tests/conf/ssl4eo_s12_simclr_2.yaml

@@ -0,0 +1,17 @@
+module:
+  _target_: torchgeo.trainers.SimCLRTask
+  model: "resnet18"
+  in_channels: 13
+  version: 2
+  layers: 3
+  hidden_dim: 8
+  output_dim: 8
+  weight_decay: 1e-4
+  memory_bank_size: 10
+
+datamodule:
+  _target_: torchgeo.datamodules.SSL4EOS12DataModule
+  root: "tests/data/ssl4eo/s12"
+  seasons: 2
+  batch_size: 2
+  num_workers: 0

tests/trainers/test_classification.py

@@ -33,7 +33,8 @@ class ClassificationTestModel(Module):
         super().__init__()
         self.conv1 = nn.Conv2d(in_channels=in_chans, out_channels=1, kernel_size=1)
         self.pool = nn.AdaptiveAvgPool2d((1, 1))
-        self.fc = nn.Linear(1, num_classes)
+        self.fc = nn.Linear(1, num_classes) if num_classes else nn.Identity()
+        self.num_features = 1
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         x = self.conv1(x)

tests/trainers/test_simclr.py

@@ -0,0 +1,154 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+
+import os
+from pathlib import Path
+from typing import Any
+
+import pytest
+import timm
+import torch
+import torchvision
+from _pytest.fixtures import SubRequest
+from _pytest.monkeypatch import MonkeyPatch
+from hydra.utils import instantiate
+from lightning.pytorch import Trainer
+from omegaconf import OmegaConf
+from torch.nn import Module
+from torchvision.models._api import WeightsEnum
+
+from torchgeo.datasets import SSL4EOS12, SeasonalContrastS2
+from torchgeo.models import get_model_weights, list_models
+from torchgeo.trainers import SimCLRTask
+
+from .test_classification import ClassificationTestModel
+
+
+def create_model(*args: Any, **kwargs: Any) -> Module:
+    return ClassificationTestModel(**kwargs)
+
+
+def load(url: str, *args: Any, **kwargs: Any) -> dict[str, Any]:
+    state_dict: dict[str, Any] = torch.load(url)
+    return state_dict
+
+
+class TestSimCLRTask:
+    @pytest.mark.parametrize(
+        "name",
+        [
+            "chesapeake_cvpr_prior_simclr",
+            "seco_simclr_1",
+            "seco_simclr_2",
+            "ssl4eo_s12_simclr_1",
+            "ssl4eo_s12_simclr_2",
+        ],
+    )
+    def test_trainer(
+        self, monkeypatch: MonkeyPatch, name: str, fast_dev_run: bool
+    ) -> None:
+        conf = OmegaConf.load(os.path.join("tests", "conf", name + ".yaml"))
+
+        if name.startswith("seco"):
+            monkeypatch.setattr(SeasonalContrastS2, "__len__", lambda self: 2)
+
+        if name.startswith("ssl4eo_s12"):
+            monkeypatch.setattr(SSL4EOS12, "__len__", lambda self: 2)
+
+        # Instantiate datamodule
+        datamodule = instantiate(conf.datamodule)
+
+        # Instantiate model
+        monkeypatch.setattr(timm, "create_model", create_model)
+        model = instantiate(conf.module)
+
+        # Instantiate trainer
+        trainer = Trainer(
+            accelerator="cpu",
+            fast_dev_run=fast_dev_run,
+            log_every_n_steps=1,
+            max_epochs=1,
+        )
+        trainer.fit(model=model, datamodule=datamodule)
+
+    def test_version_warnings(self) -> None:
+        with pytest.warns(UserWarning, match="SimCLR v1 only uses 2 layers"):
+            SimCLRTask(version=1, layers=3)
+        with pytest.warns(UserWarning, match="SimCLR v1 does not use a memory bank"):
+            SimCLRTask(version=1, memory_bank_size=10)
+        with pytest.warns(UserWarning, match=r"SimCLR v2 uses 3\+ layers"):
+            SimCLRTask(version=2, layers=2)
+        with pytest.warns(UserWarning, match="SimCLR v2 uses a memory bank"):
+            SimCLRTask(version=2, memory_bank_size=0)
+
+    @pytest.fixture(
+        params=[
+            weights for model in list_models() for weights in get_model_weights(model)
+        ]
+    )
+    def weights(self, request: SubRequest) -> WeightsEnum:
+        return request.param
+
+    @pytest.fixture
+    def mocked_weights(
+        self, tmp_path: Path, monkeypatch: MonkeyPatch, weights: WeightsEnum
+    ) -> WeightsEnum:
+        path = tmp_path / f"{weights}.pth"
+        model = timm.create_model(
+            weights.meta["model"], in_chans=weights.meta["in_chans"]
+        )
+        torch.save(model.state_dict(), path)
+        try:
+            monkeypatch.setattr(weights.value, "url", str(path))
+        except AttributeError:
+            monkeypatch.setattr(weights, "url", str(path))
+        monkeypatch.setattr(torchvision.models._api, "load_state_dict_from_url", load)
+        return weights
+
+    def test_weight_file(self, checkpoint: str) -> None:
+        model_kwargs: dict[str, Any] = {"model": "resnet18", "weights": checkpoint}
+        match = "num classes .* != num classes in pretrained model"
+        with pytest.warns(UserWarning, match=match):
+            SimCLRTask(**model_kwargs)
+
+    def test_weight_enum(self, mocked_weights: WeightsEnum) -> None:
+        model_kwargs: dict[str, Any] = {
+            "model": mocked_weights.meta["model"],
+            "weights": mocked_weights,
+            "in_channels": mocked_weights.meta["in_chans"],
+        }
+        match = "num classes .* != num classes in pretrained model"
+        with pytest.warns(UserWarning, match=match):
+            SimCLRTask(**model_kwargs)
+
+    def test_weight_str(self, mocked_weights: WeightsEnum) -> None:
+        model_kwargs: dict[str, Any] = {
+            "model": mocked_weights.meta["model"],
+            "weights": str(mocked_weights),
+            "in_channels": mocked_weights.meta["in_chans"],
+        }
+        match = "num classes .* != num classes in pretrained model"
+        with pytest.warns(UserWarning, match=match):
+            SimCLRTask(**model_kwargs)
+
+    @pytest.mark.slow
+    def test_weight_enum_download(self, weights: WeightsEnum) -> None:
+        model_kwargs: dict[str, Any] = {
+            "model": weights.meta["model"],
+            "weights": weights,
+            "in_channels": weights.meta["in_chans"],
+        }
+        match = "num classes .* != num classes in pretrained model"
+        with pytest.warns(UserWarning, match=match):
+            SimCLRTask(**model_kwargs)
+
+    @pytest.mark.slow
+    def test_weight_str_download(self, weights: WeightsEnum) -> None:
+        model_kwargs: dict[str, Any] = {
+            "model": weights.meta["model"],
+            "weights": str(weights),
+            "in_channels": weights.meta["in_chans"],
+        }
+        match = "num classes .* != num classes in pretrained model"
+        with pytest.warns(UserWarning, match=match):
+            SimCLRTask(**model_kwargs)

torchgeo/trainers/__init__.py

@@ -8,6 +8,7 @@ from .classification import ClassificationTask, MultiLabelClassificationTask
 from .detection import ObjectDetectionTask
 from .regression import PixelwiseRegressionTask, RegressionTask
 from .segmentation import SemanticSegmentationTask
+from .simclr import SimCLRTask
 
 __all__ = (
     "BYOLTask",
@@ -17,4 +18,5 @@ __all__ = (
     "PixelwiseRegressionTask",
     "RegressionTask",
     "SemanticSegmentationTask",
+    "SimCLRTask",
 )

torchgeo/trainers/simclr.py

@@ -0,0 +1,274 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+
+"""SimCLR trainer for self-supervised learning (SSL)."""
+
+import os
+import warnings
+from typing import Optional, Union, cast
+
+import kornia.augmentation as K
+import timm
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from lightly.loss import NTXentLoss
+from lightly.models.modules import SimCLRProjectionHead
+from lightning import LightningModule
+from torch import Tensor
+from torch.optim import Adam, Optimizer
+from torch.optim.lr_scheduler import CosineAnnealingLR, LinearLR, SequentialLR
+from torchvision.models._api import WeightsEnum
+
+import torchgeo.transforms as T
+
+from ..models import get_weight
+from . import utils
+
+try:
+    from torch.optim.lr_scheduler import LRScheduler
+except ImportError:
+    from torch.optim.lr_scheduler import _LRScheduler as LRScheduler
+
+
+def simclr_augmentations(size: int, weights: Tensor) -> nn.Module:
+    """Data augmentation used by SimCLR.
+
+    Args:
+        size: Size of patch to crop.
+        weights: Weight vector for grayscale computation.
+
+    Returns:
+        Data augmentation pipeline.
+    """
+    # https://github.com/google-research/simclr/blob/master/data_util.py
+    ks = size // 10 // 2 * 2 + 1
+    return K.AugmentationSequential(
+        K.RandomResizedCrop(size=(size, size), ratio=(0.75, 1.33)),
+        K.RandomHorizontalFlip(),
+        K.RandomVerticalFlip(),  # added
+        # Not appropriate for multispectral imagery, seasonal contrast used instead
+        # K.ColorJitter(brightness=0.8, contrast=0.8, saturation=0.8, hue=0.2, p=0.8)
+        T.RandomGrayscale(weights=weights, p=0.2),
+        K.RandomGaussianBlur(kernel_size=(ks, ks), sigma=(0.1, 2)),
+        data_keys=["input"],
+    )
+
+
+class SimCLRTask(LightningModule):  # type: ignore[misc]
+    """SimCLR: a simple framework for contrastive learning of visual representations.
+
+    Reference implementation:
+
+    * https://github.com/google-research/simclr
+
+    If you use this trainer in your research, please cite the following papers:
+
+    * v1: https://arxiv.org/abs/2002.05709
+    * v2: https://arxiv.org/abs/2006.10029
+
+    .. versionadded:: 0.5
+    """
+
+    def __init__(
+        self,
+        model: str = "resnet50",
+        weights: Optional[Union[WeightsEnum, str, bool]] = None,
+        in_channels: int = 3,
+        version: int = 2,
+        layers: int = 3,
+        hidden_dim: Optional[int] = None,
+        output_dim: Optional[int] = None,
+        lr: float = 4.8,
+        weight_decay: float = 1e-4,
+        temperature: float = 0.07,
+        memory_bank_size: int = 64000,
+        gather_distributed: bool = False,
+        size: int = 224,
+        grayscale_weights: Optional[Tensor] = None,
+        augmentations: Optional[nn.Module] = None,
+    ) -> None:
+        """Initialize a new SimCLRTask instance.
+
+        Args:
+            model: Name of the timm model to use.
+            weights: Initial model weights. Either a weight enum, the string
+                representation of a weight enum, True for ImageNet weights, False
+                or None for random weights, or the path to a saved model state dict.
+            in_channels: Number of input channels to model.
+            version: Version of SimCLR, 1--2.
+            layers: Number of layers in projection head (2 for v1, 3+ for v2).
+            hidden_dim: Number of hidden dimensions in projection head
+                (defaults to output dimension of model).
+            output_dim: Number of output dimensions in projection head
+                (defaults to output dimension of model).
+            lr: Learning rate
+                (0.3 x batch_size / 256 for v1, 0.3 x sqrt(batch size) for v2).
+            weight_decay: Weight decay coefficient (1e-6 for v1, 1e-4 for v2).
+            temperature: Temperature used in NT-Xent loss.
+            memory_bank_size: Size of memory bank (0 for v1, 64K for v2).
+            gather_distributed: Gather negatives from all GPUs during distributed
+                training (ignored if memory_bank_size > 0).
+            size: Size of patch to crop.
+            grayscale_weights: Weight vector for grayscale computation, see
+                :class:`~torchgeo.transforms.RandomGrayscale`. Only used when
+                ``augmentations=None``. Defaults to average of all bands.
+            augmentations: Data augmentation. Defaults to SimCLR augmentation.
+
+        Raises:
+            AssertionError: If an invalid version of SimCLR is requested.
+
+        Warns:
+            UserWarning: If hyperparameters do not match SimCLR version requested.
+        """
+        super().__init__()
+
+        # Validate hyperparameters
+        assert version in range(1, 3)
+        if version == 1:
+            if layers > 2:
+                warnings.warn("SimCLR v1 only uses 2 layers in its projection head")
+            if memory_bank_size > 0:
+                warnings.warn("SimCLR v1 does not use a memory bank")
+        elif version == 2:
+            if layers == 2:
+                warnings.warn("SimCLR v2 uses 3+ layers in its projection head")
+            if memory_bank_size == 0:
+                warnings.warn("SimCLR v2 uses a memory bank")
+
+        self.save_hyperparameters(ignore=["augmentations"])
+
+        grayscale_weights = grayscale_weights or torch.ones(in_channels)
+        self.augmentations = augmentations or simclr_augmentations(
+            size, grayscale_weights
+        )
+
+        # Create backbone
+        self.backbone = timm.create_model(
+            model, in_chans=in_channels, num_classes=0, pretrained=weights is True
+        )
+
+        # Load weights
+        if weights and weights is not True:
+            if isinstance(weights, WeightsEnum):
+                state_dict = weights.get_state_dict(progress=True)
+            elif os.path.exists(weights):
+                _, state_dict = utils.extract_backbone(weights)
+            else:
+                state_dict = get_weight(weights).get_state_dict(progress=True)
+            self.backbone = utils.load_state_dict(self.backbone, state_dict)
+
+        # Create projection head
+        input_dim = self.backbone.num_features
+        if hidden_dim is None:
+            hidden_dim = input_dim
+        if output_dim is None:
+            output_dim = input_dim
+
+        self.projection_head = SimCLRProjectionHead(
+            input_dim, hidden_dim, output_dim, layers
+        )
+
+        # Define loss function
+        self.criterion = NTXentLoss(temperature, memory_bank_size, gather_distributed)
+
+        # Initialize moving average of output
+        self.avg_output_std = 0.0
+
+        # TODO
+        # v1+: add global batch norm
+        # v2: add selective kernels, channel-wise attention mechanism
+
+    def forward(self, x: Tensor) -> tuple[Tensor, Tensor]:
+        """Forward pass of the model.
+
+        Args:
+            x: Mini-batch of images.
+
+        Returns:
+            Output from the backbone and projection head.
+        """
+        h = self.backbone(x)  # shape of batch_size x num_features
+        z = self.projection_head(h)
+        return cast(Tensor, z), cast(Tensor, h)
+
+    def training_step(self, batch: dict[str, Tensor], batch_idx: int) -> Tensor:
+        """Compute the training loss and additional metrics.
+
+        Args:
+            batch: The output of your DataLoader.
+            batch_idx: Integer displaying index of this batch.
+
+        Returns:
+            The loss tensor.
+        """
+        x = batch["image"]
+
+        in_channels = self.hparams["in_channels"]
+        assert x.size(1) == in_channels or x.size(1) == 2 * in_channels
+
+        if x.size(1) == in_channels:
+            x1 = x
+            x2 = x
+        else:
+            x1 = x[:, :in_channels]
+            x2 = x[:, in_channels:]
+
+        with torch.no_grad():
+            x1 = self.augmentations(x1)
+            x2 = self.augmentations(x2)
+
+        z1, h1 = self(x1)
+        z2, h2 = self(x2)
+
+        loss = self.criterion(z1, z2)
+
+        # Calculate the mean normalized standard deviation over features dimensions.
+        # If this is << 1 / sqrt(h1.shape[1]), then the model is not learning anything.
+        output = h1.detach()
+        output = F.normalize(output, dim=1)
+        output_std = torch.std(output, dim=0)
+        output_std = torch.mean(output_std, dim=0)
+        self.avg_output_std = 0.9 * self.avg_output_std + (1 - 0.9) * output_std.item()
+
+        self.log("train_ssl_std", self.avg_output_std)
+        self.log("train_loss", loss)
+
+        return cast(Tensor, loss)
+
+    def validation_step(self, batch: dict[str, Tensor], batch_idx: int) -> None:
+        """No-op, does nothing."""
+
+    def test_step(self, batch: dict[str, Tensor], batch_idx: int) -> None:
+        """No-op, does nothing."""
+        # TODO
+        # v2: add distillation step
+
+    def predict_step(self, batch: dict[str, Tensor], batch_idx: int) -> None:
+        """No-op, does nothing."""
+
+    def configure_optimizers(self) -> tuple[list[Optimizer], list[LRScheduler]]:
+        """Initialize the optimizer and learning rate scheduler.
+
+        Returns:
+            Optimizer and learning rate scheduler.
+        """
+        # Original paper uses LARS optimizer, but this is not defined in PyTorch
+        optimizer = Adam(
+            self.parameters(),
+            lr=self.hparams["lr"],
+            weight_decay=self.hparams["weight_decay"],
+        )
+        if self.hparams["version"] == 1:
+            warmup_epochs = 10
+        else:
+            warmup_epochs = int(self.trainer.max_epochs * 0.05)
+        lr_scheduler = SequentialLR(
+            optimizer,
+            schedulers=[
+                LinearLR(optimizer, total_iters=warmup_epochs),
+                CosineAnnealingLR(optimizer, T_max=self.trainer.max_epochs),
+            ],
+            milestones=[warmup_epochs],
+        )
+        return [optimizer], [lr_scheduler]

torchgeo/trainers/utils.py

@@ -90,6 +90,9 @@ def load_state_dict(model: Module, state_dict: "OrderedDict[str, Tensor]") -> Mo
     expected_in_channels = state_dict[input_module_key + ".weight"].shape[1]
 
     output_module_key, output_module = list(model.named_children())[-1]
+    if isinstance(output_module, nn.Identity):
+        num_classes = model.num_features
+    else:
         num_classes = output_module.out_features
     expected_num_classes = None
     if output_module_key + ".weight" in state_dict:

train.py

@@ -16,7 +16,7 @@ from lightning.pytorch.loggers import CSVLogger, TensorBoardLogger
 from omegaconf import DictConfig, OmegaConf
 
 from torchgeo.datamodules import MisconfigurationException
-from torchgeo.trainers import BYOLTask, ObjectDetectionTask
+from torchgeo.trainers import BYOLTask, ObjectDetectionTask, SimCLRTask
 
 
 def set_up_omegaconf() -> DictConfig:
@@ -95,7 +95,7 @@ def main(conf: DictConfig) -> None:
     if isinstance(task, ObjectDetectionTask):
         monitor_metric = "val_map"
         mode = "max"
-    elif isinstance(task, BYOLTask):
+    elif isinstance(task, (BYOLTask, SimCLRTask)):
         monitor_metric = "train_loss"
         mode = "min"
     else: