DataPipeline PoC

.github/workflows/ci-notebook.yml

@@ -56,10 +56,10 @@ jobs:
         # Look to see if there is a cache hit for the corresponding requirements file
         key: flash-datasets_predict
 
-    #- name: Run Notebooks
-    #  run: |
-    #    jupyter nbconvert --to script flash_notebooks/image_classification.ipynb
-    #    jupyter nbconvert --to script flash_notebooks/tabular_classification.ipynb
-    #
-    #    ipython flash_notebooks/image_classification.py
-    #    ipython flash_notebooks/tabular_classification.py
+    - name: Run Notebooks
+      run: |
+        # jupyter nbconvert --to script flash_notebooks/image_classification.ipynb
+        jupyter nbconvert --to script flash_notebooks/tabular_classification.ipynb
+
+        # ipython flash_notebooks/image_classification.py
+        ipython flash_notebooks/tabular_classification.py

.github/workflows/ci-testing.yml

@@ -60,8 +60,8 @@ jobs:
       run: |
         python --version
         pip --version
-        pip install -e . --find-links https://download.pytorch.org/whl/cpu/torch_stable.html
         pip install --requirement requirements/devel.txt --pre --upgrade --find-links https://download.pytorch.org/whl/cpu/torch_stable.html
+        pip install -e . --find-links https://download.pytorch.org/whl/cpu/torch_stable.html
         pip list
       shell: bash
 

.gitignore

@@ -143,7 +143,8 @@ flash_notebooks/*.py
 flash_notebooks/data
 MNIST*
 titanic
-coco128
 hymenoptera_data
-xsum
 imdb
+xsum
+coco128
+wmt_en_ro

README.md

@@ -254,8 +254,8 @@ download_data("https://pl-flash-data.s3.amazonaws.com/titanic.zip", 'data/')
 datamodule = TabularData.from_csv(
   "./data/titanic/titanic.csv",
   test_csv="./data/titanic/test.csv",
-  categorical_input=["Sex", "Age", "SibSp", "Parch", "Ticket", "Cabin", "Embarked"],
-  numerical_input=["Fare"],
+  cat_cols=["Sex", "Age", "SibSp", "Parch", "Ticket", "Cabin", "Embarked"],
+  num_cols=["Fare"],
   target="Survived",
   val_size=0.25,
 )

docs/source/custom_task.rst

@@ -67,33 +67,6 @@ for the prediction of diabetes disease progression. We can create this
 ``DataModule`` below, wrapping the scikit-learn `Diabetes
 dataset <https://scikit-learn.org/stable/datasets/toy_dataset.html#diabetes-dataset>`__.
 
-.. testcode::
-
-    class DiabetesPipeline(flash.core.data.TaskDataPipeline):
-        def after_uncollate(self, samples):
-            return [f"disease progression: {float(s):.2f}" for s in samples]
-
-    class DiabetesData(flash.DataModule):
-        def __init__(self, batch_size=64, num_workers=0):
-            x, y = datasets.load_diabetes(return_X_y=True)
-            x = torch.from_numpy(x).float()
-            y = torch.from_numpy(y).float().unsqueeze(1)
-            x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=.20, random_state=0)
-
-            train_ds = TensorDataset(x_train, y_train)
-            test_ds = TensorDataset(x_test, y_test)
-
-            super().__init__(
-                train_ds=train_ds,
-                test_ds=test_ds,
-                batch_size=batch_size,
-                num_workers=num_workers
-            )
-            self.num_inputs = x.shape[1]
-
-        @staticmethod
-        def default_pipeline():
-            return DiabetesPipeline()
 
 You’ll notice we added a ``DataPipeline``, which will be used when we
 call ``.predict()`` on our model. In this case we want to nicely format

docs/source/general/data.rst

@@ -7,50 +7,6 @@ Data
 DataPipeline
 ------------
 
-To make tasks work for inference, one must create a ``DataPipeline``.
-The ``flash.core.data.DataPipeline`` exposes 6 hooks to override:
-
-.. code:: python
-
-    class DataPipeline:
-        """
-        This class purpose is to facilitate the conversion of raw data to processed or batched data and back.
-        Several hooks are provided for maximum flexibility.
-
-        collate_fn:
-            - before_collate
-            - collate
-            - after_collate
-
-        uncollate_fn:
-            - before_uncollate
-            - uncollate
-            - after_uncollate
-        """
-
-        def before_collate(self, samples: Any) -> Any:
-            """Override to apply transformations to samples"""
-            return samples
-
-        def collate(self, samples: Any) -> Any:
-            """Override to convert a set of samples to a batch"""
-            if not isinstance(samples, Tensor):
-                return default_collate(samples)
-            return samples
-
-        def after_collate(self, batch: Any) -> Any:
-            """Override to apply transformations to the batch"""
-            return batch
-
-        def before_uncollate(self, batch: Any) -> Any:
-            """Override to apply transformations to the batch"""
-            return batch
-
-        def uncollate(self, batch: Any) -> ny:
-            """Override to convert a batch to a set of samples"""
-            samples = batch
-            return samples
-
-        def after_uncollate(self, samples: Any) -> Any:
-            """Override to apply transformations to samples"""
-            return samplesA
+To make tasks work for inference, one must create a ``Preprocess`` and ``PostProcess``.
+The ``flash.data.process.Preprocess`` exposes 9 hooks to override which can specifialzed for each stage using
+``train``, ``val``, ``test``, ``predict`` prefixes.

docs/source/reference/image_classification.rst

@@ -45,7 +45,7 @@ Use the :class:`~flash.vision.ImageClassifier` pretrained model for inference on
     print(predictions)
 
     # 3b. Or generate predictions with a whole folder!
-    datamodule = ImageClassificationData.from_folder(folder="data/hymenoptera_data/predict/")
+    datamodule = ImageClassificationData.from_folders(predict_folder="data/hymenoptera_data/predict/")
     predictions = Trainer().predict(model, datamodule=datamodule)
     print(predictions)
 
@@ -185,5 +185,3 @@ ImageClassificationData
 .. automethod:: flash.vision.ImageClassificationData.from_filepaths
 
 .. automethod:: flash.vision.ImageClassificationData.from_folders
-
-.. automethod:: flash.vision.ImageClassificationData.from_folder

docs/source/reference/image_embedder.rst

@@ -23,14 +23,14 @@ Use the :class:`~flash.vision.ImageEmbedder` pretrained model for inference on a
 
 .. code-block:: python
 
-	from flash.vision import ImageEmbedder
+    from flash.vision import ImageEmbedder
 
-	# Load finetuned task
-	embedder = ImageEmbedder(backbone="resnet18")
+    # Load finetuned task
+    embedder = ImageEmbedder(backbone="resnet18")
 
-	# 2. Perform inference on an image file
-	embeddings = embedder.predict("path/to/image.png")
-	print(embeddings)
+    # 2. Perform inference on an image file
+    embeddings = embedder.predict("path/to/image.png")
+    print(embeddings)
 
 Or on a random image tensor
 

docs/source/reference/object_detection.rst

@@ -41,14 +41,14 @@ Use the :class:`~flash.vision.ObjectDetector` pretrained model for inference on
 
 .. code-block:: python
 
-	from flash.vision import ObjectDetector
+    from flash.vision import ObjectDetector
 
-	# 1. Load the model
-	detector = ObjectDetector()
+    # 1. Load the model
+    detector = ObjectDetector()
 
-	# 2. Perform inference on an image file
-	predictions = detector.predict("path/to/image.png")
-	print(predictions)
+    # 2. Perform inference on an image file
+    predictions = detector.predict("path/to/image.png")
+    print(predictions)
 
 Or on a random image tensor
 

docs/source/reference/tabular_classification.rst

@@ -35,8 +35,8 @@ We can use the Flash Tabular classification task to predict the probability a pa
 We can create :class:`~flash.tabular.TabularData` from csv files using the :func:`~flash.tabular.TabularData.from_csv` method. We will pass in:
 
 * **train_csv**- csv file containing the training data converted to a Pandas DataFrame
-* **categorical_input**- a list of the names of columns that contain categorical data (strings or integers)
-* **numerical_input**- a list of the names of columns that contain numerical continuous data (floats)
+* **cat_cols**- a list of the names of columns that contain categorical data (strings or integers)
+* **num_cols**- a list of the names of columns that contain numerical continuous data (floats)
 * **target**- the name of the column we want to predict
 
 
@@ -56,8 +56,8 @@ Next, we create the :class:`~flash.tabular.TabularClassifier` task, using the Da
     datamodule = TabularData.from_csv(
         "./data/titanic/titanic.csv",
         test_csv="./data/titanic/test.csv",
-        categorical_input=["Sex", "Age", "SibSp", "Parch", "Ticket", "Cabin", "Embarked"],
-        numerical_input=["Fare"],
+        cat_cols=["Sex", "Age", "SibSp", "Parch", "Ticket", "Cabin", "Embarked"],
+        num_cols=["Fare"],
         target="Survived",
         val_size=0.25,
         )
@@ -120,8 +120,8 @@ Or you can finetune your own model and use that for prediction:
     datamodule = TabularData.from_csv(
         "my_data_file.csv",
         test_csv="./data/titanic/test.csv",
-        categorical_input=["Sex", "Age", "SibSp", "Parch", "Ticket", "Cabin", "Embarked"],
-        numerical_input=["Fare"],
+        cat_cols=["Sex", "Age", "SibSp", "Parch", "Ticket", "Cabin", "Embarked"],
+        num_cols=["Fare"],
         target="Survived",
         val_size=0.25,
     )
@@ -166,4 +166,3 @@ TabularData
 .. automethod:: flash.tabular.TabularData.from_csv
 
 .. automethod:: flash.tabular.TabularData.from_df
-

flash/__init__.py

@@ -28,12 +28,11 @@
 _PROJECT_ROOT = os.path.dirname(_PACKAGE_ROOT)
 
 from flash import tabular, text, vision  # noqa: E402
-from flash.core import data, utils  # noqa: E402
 from flash.core.classification import ClassificationTask  # noqa: E402
-from flash.core.data import DataModule  # noqa: E402
-from flash.core.data.utils import download_data  # noqa: E402
 from flash.core.model import Task  # noqa: E402
 from flash.core.trainer import Trainer  # noqa: E402
+from flash.data.data_module import DataModule  # noqa: E402
+from flash.data.utils import download_data  # noqa: E402
 
 __all__ = [
     "Task",
@@ -42,7 +41,5 @@
     "vision",
     "text",
     "tabular",
-    "data",
-    "utils",
     "download_data",
 ]

flash/core/classification.py

@@ -16,23 +16,18 @@
 import torch
 from torch import Tensor
 
-from flash.core.data import TaskDataPipeline
 from flash.core.model import Task
+from flash.data.process import Postprocess
 
 
-class ClassificationDataPipeline(TaskDataPipeline):
+class ClassificationPostprocess(Postprocess):
 
-    def before_uncollate(self, batch: Union[Tensor, tuple]) -> Tensor:
-        if isinstance(batch, tuple):
-            batch = batch[0]
-        return torch.softmax(batch, -1)
-
-    def after_uncollate(self, samples: Any) -> Any:
+    def per_sample_transform(self, samples: Any) -> Any:
         return torch.argmax(samples, -1).tolist()
 
 
 class ClassificationTask(Task):
 
-    @staticmethod
-    def default_pipeline() -> ClassificationDataPipeline:
-        return ClassificationDataPipeline()
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self._postprocess = ClassificationPostprocess()