Reviewed DataModule design+docs+types

In this MR, I try to review the use of the data module and go further in simplifications for the whole code:

• Slightly streamlined the datamodule approach
• Added documentation
• Added type annotations
• Added some TODOs for further discussion

Closes #7 (closed).

TODO:

• (@dcarron) Implement/document/type hint normaliser (ensure it is used correctly through all the models; integrate on the train script) and that train CLI works!
• (@andre.anjos) Deal with import reweight_BCEWithLogitsLoss

src/ptbench/data/shenzhen/default.py

@@ -4,19 +4,38 @@
 
 """Shenzhen datamodule for computer-aided diagnosis (default protocol)
 
-See :py:mod:`ptbench.data.shenzhen` for dataset details.
+See :py:mod:`ptbench.data.shenzhen` for more database details.
 
 This configuration:
-* raw data (default): :py:obj:`ptbench.data.shenzhen._tranforms`
-* augmentations: elastic deformation (probability = 80%)
-* output image resolution: 512x512 pixels
+
+* Raw data input (on disk):
+
+  * PNG images (black and white, encoded as color images)
+  * Variable width and height:
+
+    * widths: from 1130 to 3001 pixels
+    * heights: from 948 to 3001 pixels
+
+* Output image:
+
+  * Transforms:
+
+    * Load raw PNG with :py:mod:`PIL`
+    * Remove black borders
+    * Torch resizing(512px, 512px)
+    * Torch center cropping (512px, 512px)
+
+  * Final specifications:
+
+    * Fixed resolution: 512x512 pixels
+    * Color RGB encoding
 """
 
 import importlib.resources
 
 from ..datamodule import CachingDataModule
 from ..split import JSONDatabaseSplit
-from .raw_data_loader import raw_data_loader
+from .loader import RawDataLoader
 
 datamodule = CachingDataModule(
     database_split=JSONDatabaseSplit(
@@ -24,12 +43,5 @@ datamodule = CachingDataModule(
             "default.json.bz2"
         )
     ),
-    raw_data_loader=raw_data_loader,
-    cache_samples=False,
-    # train_sampler: typing.Optional[torch.utils.data.Sampler] = None,
-    # model_transforms = [],
-    # batch_size = 1,
-    # batch_chunk_count = 1,
-    # drop_incomplete_batch = False,
-    # parallel = -1,
+    raw_data_loader=RawDataLoader(),
 )