Change regression task to timm support (microsoft#854)

* change regression task to timm support

* add docstring about available models

* typo again

* failing test

* change name

* change name

* expose all available models

* docstring list_models

* Update torchgeo/trainers/regression.py

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

tests/conf/cowc_counting.yaml

@@ -2,6 +2,9 @@ experiment:
   task: cowc_counting
   module:
     model: resnet18
+    weights: "random"
+    num_outputs: 1
+    in_channels: 3
     learning_rate: 1e-3
     learning_rate_schedule_patience: 2
     pretrained: True

tests/conf/cyclone.yaml

@@ -2,6 +2,9 @@ experiment:
   task: "cyclone"
   module:
     model: "resnet18"
+    weights: "random"
+    num_outputs: 1
+    in_channels: 3  
     learning_rate: 1e-3
     learning_rate_schedule_patience: 2
     pretrained: False

tests/trainers/test_regression.py

@@ -5,7 +5,6 @@
 from typing import Any, Dict, Type, cast
 
 import pytest
-from _pytest.monkeypatch import MonkeyPatch
 from omegaconf import OmegaConf
 from pytorch_lightning import LightningDataModule, Trainer
 
@@ -63,22 +62,37 @@ def test_no_logger(self) -> None:
         )
         trainer.fit(model=model, datamodule=datamodule)
 
-    def test_invalid_model(self) -> None:
-        match = "module 'torchvision.models' has no attribute 'invalid_model'"
-        with pytest.raises(AttributeError, match=match):
-            RegressionTask(model="invalid_model", pretrained=False)
-
     @pytest.fixture
     def model_kwargs(self) -> Dict[Any, Any]:
-        return {"model": "resnet18", "pretrained": False}
-
-    def test_missing_attributes(
-        self, model_kwargs: Dict[Any, Any], monkeypatch: MonkeyPatch
+        return {
+            "model": "resnet18",
+            "weights": "random",
+            "num_outputs": 1,
+            "in_channels": 3,
+        }
+
+    def test_invalid_pretrained(
+        self, model_kwargs: Dict[Any, Any], checkpoint: str
     ) -> None:
-        monkeypatch.delattr(COWCCountingDataModule, "plot")
-        datamodule = COWCCountingDataModule(
-            root="tests/data/cowc_counting", batch_size=1, num_workers=0
-        )
-        model = RegressionTask(**model_kwargs)
-        trainer = Trainer(fast_dev_run=True, log_every_n_steps=1, max_epochs=1)
-        trainer.validate(model=model, datamodule=datamodule)
+        model_kwargs["weights"] = checkpoint
+        model_kwargs["model"] = "resnet50"
+        match = "Trying to load resnet18 weights into a resnet50"
+        with pytest.raises(ValueError, match=match):
+            RegressionTask(**model_kwargs)
+
+    def test_pretrained(self, model_kwargs: Dict[Any, Any], checkpoint: str) -> None:
+        model_kwargs["weights"] = checkpoint
+        with pytest.warns(UserWarning):
+            RegressionTask(**model_kwargs)
+
+    def test_invalid_model(self, model_kwargs: Dict[Any, Any]) -> None:
+        model_kwargs["model"] = "invalid_model"
+        match = "Model type 'invalid_model' is not a valid timm model."
+        with pytest.raises(ValueError, match=match):
+            RegressionTask(**model_kwargs)
+
+    def test_invalid_weights(self, model_kwargs: Dict[Any, Any]) -> None:
+        model_kwargs["weights"] = "invalid_weights"
+        match = "Weight type 'invalid_weights' is not valid."
+        with pytest.raises(ValueError, match=match):
+            RegressionTask(**model_kwargs)

torchgeo/trainers/regression.py

@@ -3,21 +3,21 @@
 
 """Regression tasks."""
 
+import os
 from typing import Any, Dict, cast
 
 import matplotlib.pyplot as plt
 import pytorch_lightning as pl
+import timm
 import torch
-import torch.nn as nn
 import torch.nn.functional as F
-import torchvision
-from packaging.version import parse
 from torch import Tensor
 from torch.nn.modules import Conv2d, Linear
 from torch.optim.lr_scheduler import ReduceLROnPlateau
 from torchmetrics import MeanAbsoluteError, MeanSquaredError, MetricCollection
 
 from ..datasets.utils import unbind_samples
+from . import utils
 
 # https://github.com/pytorch/pytorch/issues/60979
 # https://github.com/pytorch/pytorch/pull/61045
@@ -26,36 +26,74 @@
 
 
 class RegressionTask(pl.LightningModule):
-    """LightningModule for training models on regression datasets."""
+    """LightningModule for training models on regression datasets.
+
+    Supports any available `Timm model
+    <https://rwightman.github.io/pytorch-image-models/>`_
+    as an architecture choice. To see a list of available
+    models, you can do:
+
+    .. code-block:: python
+
+        import timm
+        print(timm.list_models())
+    """
 
     def config_task(self) -> None:
         """Configures the task based on kwargs parameters."""
+        in_channels = self.hyperparams["in_channels"]
         model = self.hyperparams["model"]
-        pretrained = self.hyperparams["pretrained"]
-
-        if parse(torchvision.__version__) >= parse("0.13"):
-            if pretrained:
-                kwargs = {
-                    "weights": getattr(
-                        torchvision.models, f"ResNet{model[6:]}_Weights"
-                    ).DEFAULT
-                }
+
+        imagenet_pretrained = False
+        custom_pretrained = False
+        if self.hyperparams["weights"] and not os.path.exists(
+            self.hyperparams["weights"]
+        ):
+            if self.hyperparams["weights"] not in ["imagenet", "random"]:
+                raise ValueError(
+                    f"Weight type '{self.hyperparams['weights']}' is not valid."
+                )
             else:
-                kwargs = {"weights": None}
+                imagenet_pretrained = self.hyperparams["weights"] == "imagenet"
+            custom_pretrained = False
+        else:
+            custom_pretrained = True
+
+        # Create the model
+        valid_models = timm.list_models(pretrained=imagenet_pretrained)
+        if model in valid_models:
+            self.model = timm.create_model(
+                model,
+                num_classes=self.hyperparams["num_outputs"],
+                in_chans=in_channels,
+                pretrained=imagenet_pretrained,
+            )
         else:
-            kwargs = {"pretrained": pretrained}
+            raise ValueError(f"Model type '{model}' is not a valid timm model.")
 
-        self.model = getattr(torchvision.models, model)(**kwargs)
-        in_features = self.model.fc.in_features
-        self.model.fc = nn.Linear(in_features, out_features=1)
+        if custom_pretrained:
+            name, state_dict = utils.extract_encoder(self.hyperparams["weights"])
+
+            if self.hyperparams["model"] != name:
+                raise ValueError(
+                    f"Trying to load {name} weights into a "
+                    f"{self.hyperparams['model']}"
+                )
+            self.model = utils.load_state_dict(self.model, state_dict)
 
     def __init__(self, **kwargs: Any) -> None:
         """Initialize a new LightningModule for training simple regression models.
 
         Keyword Args:
-            model: Name of the model to use
-            learning_rate: Initial learning rate to use in the optimizer
-            learning_rate_schedule_patience: Patience parameter for the LR scheduler
+            model: Name of the timm model to use
+            weights: Either "random" or "imagenet"
+            num_outputs: Number of prediction outputs
+            in_channels: Number of input channels to model
+            learning_rate: Learning rate for optimizer
+            learning_rate_schedule_patience: Patience for learning rate scheduler
+
+        .. versionchanged:: 0.4
+            Change regression model support from torchvision.models to timm
         """
         super().__init__()