Decouple DataModules from Models - GAN (#206)

* 🎨 decouple dms from gan

* ✅ update tests

* ✅ update tests

* 💄 style

* 🐛 Gan now has required args

* 🐛 use argparse not hyperopt parser

pl_bolts/models/gans/basic/basic_gan_module.py

@@ -5,20 +5,20 @@
 import torch
 from torch.nn import functional as F
 
-from pl_bolts.datamodules import MNISTDataModule
 from pl_bolts.models.gans.basic.components import Generator, Discriminator
 
 
 class GAN(pl.LightningModule):
 
-    def __init__(self,
-                 datamodule: pl.LightningDataModule = None,
-                 latent_dim: int = 32,
-                 batch_size: int = 100,
-                 learning_rate: float = 0.0002,
-                 data_dir: str = '',
-                 num_workers: int = 8,
-                 **kwargs):
+    def __init__(
+        self,
+        input_channels: int,
+        input_height: int,
+        input_width: int,
+        latent_dim: int = 32,
+        learning_rate: float = 0.0002,
+        **kwargs
+    ):
         """
         Vanilla GAN implementation.
 
@@ -53,24 +53,12 @@ def __init__(self,
 
         # makes self.hparams under the hood and saves to ckpt
         self.save_hyperparameters()
-
-        self._set_default_datamodule(datamodule)
+        self.img_dim = (input_channels, input_height, input_width)
 
         # networks
         self.generator = self.init_generator(self.img_dim)
         self.discriminator = self.init_discriminator(self.img_dim)
 
-    def _set_default_datamodule(self, datamodule):
-        # link default data
-        if datamodule is None:
-            datamodule = MNISTDataModule(
-                data_dir=self.hparams.data_dir,
-                num_workers=self.hparams.num_workers,
-                normalize=True
-            )
-        self.datamodule = datamodule
-        self.img_dim = self.datamodule.size()
-
     def init_generator(self, img_dim):
         generator = Generator(latent_dim=self.hparams.latent_dim, img_shape=img_dim)
         return generator
@@ -179,44 +167,40 @@ def add_model_specific_args(parent_parser):
                             help="adam: decay of first order momentum of gradient")
         parser.add_argument('--latent_dim', type=int, default=100,
                             help="generator embedding dim")
-        parser.add_argument('--batch_size', type=int, default=64, help="size of the batches")
-        parser.add_argument('--num_workers', type=int, default=8, help="num dataloader workers")
-        parser.add_argument('--data_dir', type=str, default=os.getcwd())
-
         return parser
 
 
-def cli_main():
+def cli_main(args=None):
     from pl_bolts.callbacks import LatentDimInterpolator, TensorboardGenerativeModelImageSampler
-    from pl_bolts.datamodules import STL10DataModule, ImagenetDataModule
+    from pl_bolts.datamodules import CIFAR10DataModule, ImagenetDataModule, MNISTDataModule, STL10DataModule
 
     pl.seed_everything(1234)
 
     parser = ArgumentParser()
-    parser.add_argument('--dataset', type=str, default='mnist', help='mnist, stl10, imagenet2012')
-
+    parser.add_argument("--dataset", default="mnist", type=str, help="mnist, cifar10, stl10, imagenet")
+    script_args, _ = parser.parse_known_args(args)
+
+    if script_args.dataset == "mnist":
+        dm_cls = MNISTDataModule
+    elif script_args.dataset == "cifar10":
+        dm_cls = CIFAR10DataModule
+    elif script_args.dataset == "stl10":
+        dm_cls = STL10DataModule
+    elif script_args.dataset == "imagenet":
+        dm_cls = ImagenetDataModule
+
+    parser = dm_cls.add_argparse_args(parser)
     parser = pl.Trainer.add_argparse_args(parser)
     parser = GAN.add_model_specific_args(parser)
-    parser = ImagenetDataModule.add_argparse_args(parser)
-    args = parser.parse_args()
-
-    # default is mnist
-    datamodule = None
-    if args.dataset == 'imagenet2012':
-        datamodule = ImagenetDataModule.from_argparse_args(args)
-    elif args.dataset == 'stl10':
-        datamodule = STL10DataModule.from_argparse_args(args)
-
-    gan = GAN(**vars(args), datamodule=datamodule)
-    callbacks = [TensorboardGenerativeModelImageSampler(), LatentDimInterpolator()]
+    args = parser.parse_args(args)
 
-    trainer = pl.Trainer.from_argparse_args(
-        args,
-        callbacks=callbacks,
-        progress_bar_refresh_rate=10
-    )
-    trainer.fit(gan)
+    dm = dm_cls.from_argparse_args(args)
+    model = GAN(*dm.size(), **vars(args))
+    callbacks = [TensorboardGenerativeModelImageSampler(), LatentDimInterpolator(interpolate_epoch_interval=5)]
+    trainer = pl.Trainer.from_argparse_args(args, callbacks=callbacks, progress_bar_refresh_rate=20)
+    trainer.fit(model, dm)
+    return dm, model, trainer
 
 
 if __name__ == '__main__':
-    cli_main()
+    dm, model, trainer = cli_main()

tests/callbacks/test_variational_callbacks.py

@@ -12,7 +12,7 @@ def __init__(self):
             self.global_step = 1
             self.logger = DummyLogger()
 
-    model = GAN()
+    model = GAN(3, 28, 28)
     cb = LatentDimInterpolator(interpolate_epoch_interval=2)
 
     cb.on_epoch_end(FakeTrainer(), model)

tests/models/test_executable_scripts.py

@@ -3,14 +3,23 @@
 import pytest
 
 
-@pytest.mark.parametrize('cli_args', ['--max_epochs 1'
-                                      ' --limit_train_batches 3'
-                                      ' --limit_val_batches 3'
-                                      ' --batch_size 3'])
-def test_cli_basic_gan(cli_args):
+@pytest.mark.parametrize(
+    "dataset_name", [
+        pytest.param('mnist', id="mnist"),
+        pytest.param('cifar10', id="cifar10")
+    ]
+)
+def test_cli_basic_gan(dataset_name):
     from pl_bolts.models.gans.basic.basic_gan_module import cli_main
 
-    cli_args = cli_args.split(' ') if cli_args else []
+    cli_args = f"""
+        --dataset {dataset_name}
+        --max_epochs 1
+        --limit_train_batches 3
+        --limit_val_batches 3
+        --batch_size 3
+    """.strip().split()
+
     with mock.patch("argparse._sys.argv", ["any.py"] + cli_args):
         cli_main()
 

tests/models/test_gans.py

@@ -1,13 +1,19 @@
+import pytest
 import pytorch_lightning as pl
 from pytorch_lightning import seed_everything
 
+from pl_bolts.datamodules import MNISTDataModule, CIFAR10DataModule
 from pl_bolts.models.gans import GAN
 
 
-def test_gan(tmpdir):
+@pytest.mark.parametrize(
+    "dm_cls", [pytest.param(MNISTDataModule, id="mnist"), pytest.param(CIFAR10DataModule, id="cifar10")]
+)
+def test_gan(tmpdir, dm_cls):
     seed_everything()
 
-    model = GAN(data_dir=tmpdir)
+    dm = dm_cls()
+    model = GAN(*dm.size())
     trainer = pl.Trainer(fast_dev_run=True, default_root_dir=tmpdir)
-    trainer.fit(model)
-    trainer.test(model)
+    trainer.fit(model, dm)
+    trainer.test(datamodule=dm)