Refactor preprocess_cls to preprocess, add Serializer, add DataPipelineState

.gitignore

@@ -1,3 +1,4 @@
+.DS_Store
 .lock
 lightning_logs
 
@@ -149,3 +150,4 @@ xsum
 coco128
 wmt_en_ro
 kinetics
+movie_posters

docs/source/custom_task.rst

@@ -115,13 +115,13 @@ We will define a custom ``NumpyDataModule`` class subclassing :class:`~flash.dat
 This ``NumpyDataModule`` class will provide a ``from_xy_dataset`` helper ``classmethod`` to instantiate
 :class:`~flash.data.data_module.DataModule` from x, y numpy arrays.
 
-Here is how it would look like:
+Here is how it would look:
 
 Example::
 
     x, y = ...
-    preprocess_cls = ...
-    datamodule = NumpyDataModule.from_xy_dataset(x, y, preprocess_cls)
+    preprocess = ...
+    datamodule = NumpyDataModule.from_xy_dataset(x, y, preprocess)
 
 Here is the ``NumpyDataModule`` implementation:
 
@@ -140,12 +140,12 @@ Example::
             cls,
             x: ND,
             y: ND,
-            preprocess_cls: Preprocess = NumpyPreprocess,
+            preprocess: Preprocess = None,
             batch_size: int = 64,
             num_workers: int = 0
         ):
 
-            preprocess = preprocess_cls()
+            preprocess = preprocess or NumpyPreprocess()
 
             x_train, x_test, y_train, y_test = train_test_split(
                 x, y, test_size=.20, random_state=0)
@@ -180,7 +180,7 @@ It allows the user much more granular control over their data processing flow.
 
 .. note::
 
-    Why introducing :class:`~flash.data.process.Preprocess` ?
+    Why introduce :class:`~flash.data.process.Preprocess` ?
 
     The :class:`~flash.data.process.Preprocess` object reduces the engineering overhead to make inference on raw data or
     to deploy the model in production environnement compared to traditional

docs/source/general/data.rst

@@ -29,7 +29,9 @@ Here are common terms you need to be familiar with:
         The :class:`~flash.data.process.Preprocess` hooks covers from data-loading to model forwarding.
    * - :class:`~flash.data.process.Postprocess`
      - The :class:`~flash.data.process.Postprocess` provides a simple hook-based API to encapsulate your post-processing logic.
-        The :class:`~flash.data.process.Postprocess` hooks covers from model outputs to predictions export.
+        The :class:`~flash.data.process.Postprocess` hooks cover from model outputs to predictions export.
+   * - :class:`~flash.data.process.Serializer`
+     - The :class:`~flash.data.process.Serializer` provides a single ``serialize`` method that is used to convert model outputs (after the :class:`~flash.data.process.Postprocess`) to the desired output format during prediction.
 
 *******************************************
 How to use out-of-the-box flashdatamodules
@@ -49,7 +51,9 @@ However, after model training, it requires a lot of engineering overhead to make
 Usually, extra processing logic should be added to bridge the gap between training data and raw data.
 
 The :class:`~flash.data.process.Preprocess` and :class:`~flash.data.process.Postprocess` classes can be used to
-store the data as well as the preprocessing and postprocessing transforms.
+store the data as well as the preprocessing and postprocessing transforms. The :class:`~flash.data.process.Serializer`
+class provides the logic for converting :class:`~flash.data.process.Postprocess` outputs to the desired predict format
+(e.g. classes, labels, probabilites, etc.).
 
 By providing a series of hooks that can be overridden with custom data processing logic,
 the user has much more granular control over their data processing flow.
@@ -122,7 +126,7 @@ Example::
         train_folder="data/hymenoptera_data/train/",
         val_folder="data/hymenoptera_data/val/",
         test_folder="data/hymenoptera_data/test/",
-        preprocess_cls=CustomImageClassificationPreprocess
+        preprocess=CustomImageClassificationPreprocess(),
     )
 
 
@@ -157,7 +161,7 @@ Example::
         val_folder="./data/val",
         test_folder="./data/test",
         predict_folder="./data/predict",
-        preprocess=preprocess
+        preprocess=preprocess,
     )
 
     model = ImageClassifier(...)
@@ -190,6 +194,7 @@ Example::
             **kwargs
         ):
 
+            # Set a custom ``Preprocess`` if none was provided
             preprocess = preprocess or cls.preprocess_cls()
 
             # {stage}_load_data_input will be given to your
@@ -291,6 +296,18 @@ ___________
     :members:
 
 
+----------
+
+.. _serializer:
+
+Serializer
+___________
+
+
+.. autoclass:: flash.data.process.Serializer
+    :members:
+
+
 ----------
 
 .. _datapipeline:
@@ -414,16 +431,18 @@ Example::
     predictions = lightning_module(data)
 
 
-Postprocess
-___________
+Postprocess and Serializer
+__________________________
 
 
-Once the predictions have been generated by the Flash :class:`~flash.core.model.Task`.
-The Flash :class:`~flash.data.data_pipeline.DataPipeline` will behind the scenes execute the :class:`~flash.data.process.Postprocess` hooks.
+Once the predictions have been generated by the Flash :class:`~flash.core.model.Task`, the Flash
+:class:`~flash.data.data_pipeline.DataPipeline` will execute the :class:`~flash.data.process.Postprocess` hooks and the
+:class:`~flash.data.process.Serializer` behind the scenes.
 
-First, the ``per_batch_transform`` hooks will be applied on the batch predictions.
-Then the ``uncollate`` will split the batch into individual predictions.
-Finally, the ``per_sample_transform`` will be applied on each prediction.
+First, the :meth:`~flash.data.process.Postprocess.per_batch_transform` hooks will be applied on the batch predictions.
+Then, the :meth:`~flash.data.process.Postprocess.uncollate` will split the batch into individual predictions.
+Next, the :meth:`~flash.data.process.Postprocess.per_sample_transform` will be applied on each prediction.
+Finally, the :meth:`~flash.data.process.Serializer.serialize` method will be called to serialize the predictions.
 
 .. note:: The transform can be applied either on device or ``CPU``.
 
@@ -438,7 +457,9 @@ Example::
 
         samples = uncollate(batch)
 
-        return [per_sample_transform(sample) for sample in samples]
+        samples = [per_sample_transform(sample) for sample in samples]
+        # only if serializers are enabled.
+        return [serialize(sample) for sample in samples]
 
     predictions = lightning_module(data)
     return uncollate_fn(predictions)

docs/source/general/predictions.rst

@@ -12,7 +12,7 @@ Predict on a single sample of data
 
 You can pass in a sample of data (image file path, a string of text, etc) to the :func:`~flash.core.model.Task.predict` method.
 
-	
+
 .. code-block:: python
 
 	from flash import Trainer
@@ -51,5 +51,3 @@ Predict on a csv file
 	# 3. Generate predictions from a csv file! Who would survive?
 	predictions = model.predict("data/titanic/titanic.csv")
 	print(predictions)
-
-

docs/source/general/training.rst

@@ -48,7 +48,7 @@ Flash tasks supports many advanced training functionalities out-of-the-box, such
 
     # train on 1 GPU
     flash.Trainer(gpus=1)
-    
+
 * Training on multiple GPUs
 
 .. code-block:: python
@@ -60,7 +60,7 @@ Flash tasks supports many advanced training functionalities out-of-the-box, such
 
     # train on gpu 1, 3, 5 (3 gpus total)
     flash.Trainer(gpus=[1, 3, 5])
-    
+
 * Using mixed precision training
 
 .. code-block:: python

docs/source/index.rst

@@ -22,6 +22,7 @@ Lightning Flash
    reference/task
    reference/image_classification
    reference/image_embedder
+   reference/multi_label_classification
    reference/summarization
    reference/text_classification
    reference/tabular_classification

docs/source/quickstart.rst

@@ -15,7 +15,7 @@ For getting started with Deep Learning
 
 Easy to learn
 ^^^^^^^^^^^^^
-If you are just getting started with deep learning, Flash offers common deep learning tasks you can use out-of-the-box in a few lines of code, no math, fancy nn.Modules or research experience required! 
+If you are just getting started with deep learning, Flash offers common deep learning tasks you can use out-of-the-box in a few lines of code, no math, fancy nn.Modules or research experience required!
 
 Easy to scale
 ^^^^^^^^^^^^^
@@ -70,7 +70,7 @@ You can install flash using pip or conda:
 Tasks
 =====
 
-Flash is comprised of a collection of Tasks. The Flash tasks are laser-focused objects designed to solve a well-defined type of problem, using state-of-the-art methods. 
+Flash is comprised of a collection of Tasks. The Flash tasks are laser-focused objects designed to solve a well-defined type of problem, using state-of-the-art methods.
 
 The Flash tasks contain all the relevant information to solve the task at hand- the number of class labels you want to predict, number of columns in your dataset, as well as details on the model architecture used such as loss function, optimizers, etc.