Reviewing GAN basics, VisionDataModule, MNISTDataModule, CIFAR10DataM…

…odule (#843)

pl_bolts/datamodules/cifar10_datamodule.py

@@ -15,7 +15,6 @@
     CIFAR10 = None
 
 
-@under_review()
 class CIFAR10DataModule(VisionDataModule):
     """
     .. figure:: https://3qeqpr26caki16dnhd19sv6by6v-wpengine.netdna-ssl.com/wp-content/uploads/2019/01/

pl_bolts/datamodules/mnist_datamodule.py

@@ -3,7 +3,6 @@
 from pl_bolts.datamodules.vision_datamodule import VisionDataModule
 from pl_bolts.datasets import MNIST
 from pl_bolts.utils import _TORCHVISION_AVAILABLE
-from pl_bolts.utils.stability import under_review
 from pl_bolts.utils.warnings import warn_missing_pkg
 
 if _TORCHVISION_AVAILABLE:
@@ -12,7 +11,6 @@
     warn_missing_pkg("torchvision")
 
 
-@under_review()
 class MNISTDataModule(VisionDataModule):
     """
     .. figure:: https://miro.medium.com/max/744/1*AO2rIhzRYzFVQlFLx9DM9A.png

pl_bolts/datamodules/vision_datamodule.py

@@ -6,10 +6,7 @@
 from pytorch_lightning import LightningDataModule
 from torch.utils.data import DataLoader, Dataset, random_split
 
-from pl_bolts.utils.stability import under_review
 
-
-@under_review()
 class VisionDataModule(LightningDataModule):
 
     EXTRA_ARGS: dict = {}
@@ -30,6 +27,9 @@ def __init__(
         shuffle: bool = True,
         pin_memory: bool = True,
         drop_last: bool = False,
+        train_transforms: Optional[Callable] = None,
+        val_transforms: Optional[Callable] = None,
+        test_transforms: Optional[Callable] = None,
         *args: Any,
         **kwargs: Any,
     ) -> None:
@@ -58,6 +58,36 @@ def __init__(
         self.shuffle = shuffle
         self.pin_memory = pin_memory
         self.drop_last = drop_last
+        self._train_transforms = train_transforms
+        self._val_transforms = val_transforms
+        self._test_transforms = test_transforms
+
+    @property
+    def train_transforms(self) -> Optional[Callable[..., Any]]:
+        """Optional transforms (or collection of transforms) you can apply to train dataset."""
+        return self._train_transforms
+
+    @train_transforms.setter
+    def train_transforms(self, t: Callable) -> None:
+        self._train_transforms = t
+
+    @property
+    def val_transforms(self) -> Optional[Callable[..., Any]]:
+        """Optional transforms (or collection of transforms) you can apply to validation dataset."""
+        return self._val_transforms
+
+    @val_transforms.setter
+    def val_transforms(self, t: Callable) -> None:
+        self._val_transforms = t
+
+    @property
+    def test_transforms(self) -> Optional[Callable[..., Any]]:
+        """Optional transforms (or collection of transforms) you can apply to test dataset."""
+        return self._test_transforms
+
+    @test_transforms.setter
+    def test_transforms(self, t: Callable) -> None:
+        self._test_transforms = t
 
     def prepare_data(self, *args: Any, **kwargs: Any) -> None:
         """Saves files to data_dir."""

pl_bolts/models/gans/basic/basic_gan_module.py

@@ -2,13 +2,12 @@
 
 import torch
 from pytorch_lightning import LightningModule, Trainer, seed_everything
+from pytorch_lightning.callbacks.progress import TQDMProgressBar
 from torch.nn import functional as F
 
 from pl_bolts.models.gans.basic.components import Discriminator, Generator
-from pl_bolts.utils.stability import under_review
 
 
-@under_review()
 class GAN(LightningModule):
     """Vanilla GAN implementation.
 
@@ -22,7 +21,7 @@ class GAN(LightningModule):
     Example CLI::
 
         # mnist
-        python  basic_gan_module.py --gpus 1
+        python basic_gan_module.py --gpus 1
 
         # imagenet
         python  basic_gan_module.py --gpus 1 --dataset 'imagenet2012'
@@ -166,7 +165,6 @@ def add_model_specific_args(parent_parser):
         return parser
 
 
-@under_review()
 def cli_main(args=None):
     from pl_bolts.callbacks import LatentDimInterpolator, TensorboardGenerativeModelImageSampler
     from pl_bolts.datamodules import CIFAR10DataModule, ImagenetDataModule, MNISTDataModule, STL10DataModule
@@ -193,8 +191,12 @@ def cli_main(args=None):
 
     dm = dm_cls.from_argparse_args(args)
     model = GAN(*dm.size(), **vars(args))
-    callbacks = [TensorboardGenerativeModelImageSampler(), LatentDimInterpolator(interpolate_epoch_interval=5)]
-    trainer = Trainer.from_argparse_args(args, callbacks=callbacks, progress_bar_refresh_rate=20)
+    callbacks = [
+        TensorboardGenerativeModelImageSampler(),
+        LatentDimInterpolator(interpolate_epoch_interval=5),
+        TQDMProgressBar(refresh_rate=20),
+    ]
+    trainer = Trainer.from_argparse_args(args, callbacks=callbacks)
     trainer.fit(model, datamodule=dm)
     return dm, model, trainer
 

pl_bolts/models/gans/basic/components.py

@@ -3,10 +3,7 @@
 from torch import nn
 from torch.nn import functional as F
 
-from pl_bolts.utils.stability import under_review
 
-
-@under_review()
 class Generator(nn.Module):
     def __init__(self, latent_dim, img_shape, hidden_dim=256):
         super().__init__()
@@ -27,7 +24,6 @@ def forward(self, z):
         return img
 
 
-@under_review()
 class Discriminator(nn.Module):
     def __init__(self, img_shape, hidden_dim=1024):
         super().__init__()

tests/datamodules/test_datamodules.py

@@ -72,17 +72,17 @@ def test_cityscapes_datamodule(datadir):
 
 
 @pytest.mark.parametrize("val_split, train_len", [(0.2, 48_000), (5_000, 55_000)])
-def test_vision_data_module(datadir, val_split, train_len):
+def test_vision_data_module(datadir, val_split, catch_warnings, train_len):
     dm = _create_dm(MNISTDataModule, datadir, val_split=val_split)
     assert len(dm.dataset_train) == train_len
 
 
 @pytest.mark.parametrize("dm_cls", [BinaryMNISTDataModule, CIFAR10DataModule, FashionMNISTDataModule, MNISTDataModule])
-def test_data_modules(datadir, dm_cls):
+def test_data_modules(datadir, catch_warnings, dm_cls):
     dm = _create_dm(dm_cls, datadir)
     loader = dm.train_dataloader()
     img, _ = next(iter(loader))
-    assert img.size() == torch.Size([2, *dm.size()])
+    assert img.size() == torch.Size([2, *dm.dims])
 
 
 def _create_dm(dm_cls, datadir, **kwargs):

tests/models/test_gans.py → tests/models/gans/integration/test_gans.py

@@ -1,5 +1,8 @@
+import warnings
+
 import pytest
 from pytorch_lightning import Trainer, seed_everything
+from pytorch_lightning.utilities.warnings import PossibleUserWarning
 from torch.utils.data.dataloader import DataLoader
 from torchvision import transforms as transform_lib
 
@@ -10,17 +13,26 @@
 
 @pytest.mark.parametrize(
     "dm_cls",
-    [
-        pytest.param(MNISTDataModule, id="mnist"),
-        pytest.param(CIFAR10DataModule, id="cifar10"),
-    ],
+    [pytest.param(MNISTDataModule, id="mnist"), pytest.param(CIFAR10DataModule, id="cifar10")],
 )
-def test_gan(tmpdir, datadir, dm_cls):
-    seed_everything()
-
+def test_gan(tmpdir, datadir, catch_warnings, dm_cls):
+    # Validation loop for GANs is not well defined!
+    warnings.filterwarnings(
+        "ignore",
+        message="You passed in a `val_dataloader` but have no `validation_step`. Skipping val loop.",
+        category=UserWarning,
+    )
+    warnings.filterwarnings(
+        "ignore",
+        message="The dataloader, train_dataloader, does not have many workers which may be a bottleneck",
+        category=PossibleUserWarning,
+    )
+    seed_everything(1234)
     dm = dm_cls(data_dir=datadir, num_workers=0)
-    model = GAN(*dm.size())
-    trainer = Trainer(fast_dev_run=True, default_root_dir=tmpdir)
+    model = GAN(*dm.dims)
+    trainer = Trainer(
+        fast_dev_run=True, default_root_dir=tmpdir, max_epochs=-1, accelerator="auto", log_every_n_steps=1
+    )
     trainer.fit(model, datamodule=dm)
 
 

tests/models/gans/unit/test_basic_components.py

@@ -0,0 +1,38 @@
+import pytest
+import torch
+from pytorch_lightning import seed_everything
+
+from pl_bolts.models.gans.basic.components import Discriminator, Generator
+
+
+@pytest.mark.parametrize(
+    "latent_dim, img_shape",
+    [
+        pytest.param(100, (3, 28, 28), id="100-multichannel"),
+        pytest.param(100, (1, 28, 28), id="100-singlechannel"),
+    ],
+)
+def test_generator(catch_warnings, latent_dim, img_shape):
+    batch_dim = 10
+    seed_everything()
+    generator = Generator(latent_dim=latent_dim, img_shape=img_shape)
+    noise = torch.randn(batch_dim, latent_dim)
+    samples = generator(noise)
+    assert samples.shape == (batch_dim, *img_shape)
+
+
+@pytest.mark.parametrize(
+    "img_shape",
+    [
+        pytest.param((3, 28, 28), id="discriminator-multichannel"),
+        pytest.param((1, 28, 28), id="discriminator-singlechannel"),
+    ],
+)
+def test_discriminator(catch_warnings, img_shape):
+    batch_dim = 10
+    seed_everything()
+    discriminator = Discriminator(img_shape=img_shape)
+    samples = torch.randn(batch_dim, *img_shape)
+    real_or_fake = discriminator(samples)
+    assert real_or_fake.shape == (batch_dim, 1)
+    assert (torch.clamp(real_or_fake.clone(), 0, 1) == real_or_fake).all()