Data Pipeline V2: Refactor Deserializer into Serve Input

CHANGELOG.md

@@ -32,6 +32,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Changed the `SpeechRecognition` task to use `AutoModelForCTC` rather than just `Wav2Vec2ForCTC` ([#874](https://github.com/PyTorchLightning/lightning-flash/pull/874))
 
+- Changed the `Deserializer` to subclass `ServeInput` ([#1013](https://github.com/PyTorchLightning/lightning-flash/pull/1013))
+
 - Added `Output` suffix to `Preds`, `FiftyOneDetectionLabels`, `SegmentationLabels`, `FiftyOneDetectionLabels`, `DetectionLabels`, `Classes`, `FiftyOneLabels`, `Labels`, `Logits`, `Probabilities` ([#1011](https://github.com/PyTorchLightning/lightning-flash/pull/1011))
 
 ### Deprecated

flash/audio/speech_recognition/data.py

@@ -43,12 +43,12 @@
 
 
 class SpeechRecognitionDeserializer(Deserializer):
-    def __init__(self, sampling_rate: int = 16000):
-        super().__init__()
-
+    @requires("audio")
+    def __init__(self, sampling_rate: int = 16000, **kwargs):
+        super().__init__(**kwargs)
         self.sampling_rate = sampling_rate
 
-    def deserialize(self, sample: Any) -> Dict:
+    def serve_load_sample(self, sample: Any) -> Dict:
         encoded_with_padding = (sample + "===").encode("ascii")
         audio = base64.b64decode(encoded_with_padding)
         buffer = io.BytesIO(audio)

flash/core/data/batch.py

@@ -42,7 +42,8 @@ def __init__(
         self._per_sample_transform_context = CurrentFuncContext("per_sample_transform", input_transform)
 
     def forward(self, sample: str):
-        sample = self.deserializer(sample)
+
+        sample = self.deserializer.deserialize(sample)
 
         with self._current_stage_context:
             with self._per_sample_transform_context:

flash/core/data/data_pipeline.py

@@ -197,6 +197,8 @@ def _resolve_function_hierarchy(
             prefixes += ["test"]
         elif stage == RunningStage.PREDICTING:
             prefixes += ["predict"]
+        elif stage == RunningStage.SERVING:
+            prefixes += ["serve"]
 
         prefixes += [None]
 

flash/core/data/io/input_base.py

@@ -15,8 +15,9 @@
 import os
 import sys
 from copy import copy, deepcopy
-from typing import Any, cast, Dict, Iterable, MutableMapping, Optional, Sequence, Tuple, Union
+from typing import Any, cast, Dict, Iterable, List, MutableMapping, Optional, Sequence, Tuple, Union
 
+from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from torch.utils.data import Dataset
 
 import flash
@@ -232,3 +233,26 @@ def __iter__(self):
 
     def __next__(self) -> Any:
         return self._call_load_sample(next(self.data_iter))
+
+
+class ServeInput(Input):
+    def __init__(
+        self,
+        data_pipeline_state: Optional["flash.core.data.data_pipeline.DataPipelineState"] = None,
+    ) -> None:
+        if hasattr(self, "serve_load_data"):
+            raise MisconfigurationException("`serve_load_data` shouldn't be implemented.")
+
+        super().__init__(RunningStage.SERVING, data_pipeline_state=data_pipeline_state)
+
+    def serve_load_sample(self, sample: Any) -> List[Any]:
+        raise NotImplementedError
+
+    def deserialize(self, sample: Any) -> Any:
+        return self._call_load_sample(sample)
+
+    def example_input(self) -> str:
+        raise NotImplementedError
+
+    def __bool__(self):
+        return True

flash/core/data/process.py

@@ -12,30 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import functools
-from abc import abstractmethod
 from typing import Any, Mapping
 from warnings import warn
 
 from deprecate import deprecated
 
 import flash
+from flash.core.data.io.input_base import ServeInput as Deserializer
 from flash.core.data.io.output import Output
-from flash.core.data.properties import Properties
-
-
-class Deserializer(Properties):
-    """Deserializer."""
-
-    def deserialize(self, sample: Any) -> Any:  # TODO: Output must be a tensor???
-        raise NotImplementedError
-
-    @property
-    @abstractmethod
-    def example_input(self) -> str:
-        raise NotImplementedError
-
-    def __call__(self, sample: Any) -> Any:
-        return self.deserialize(sample)
 
 
 class DeserializerMapping(Deserializer):

flash/core/model.py

@@ -809,7 +809,13 @@ def build_data_pipeline(
         if deserializer is None or type(deserializer) is Deserializer:
             deserializer = getattr(input_transform, "deserializer", deserializer)
 
-        data_pipeline = DataPipeline(input, input_transform, output_transform, deserializer, output)
+        data_pipeline = DataPipeline(
+            input=input,
+            input_transform=input_transform,
+            output_transform=output_transform,
+            deserializer=deserializer,
+            output=output,
+        )
         self._data_pipeline_state = self._data_pipeline_state or DataPipelineState()
         self.attach_data_pipeline_state(self._data_pipeline_state)
         self._data_pipeline_state = data_pipeline.initialize(self._data_pipeline_state)

flash/image/data.py

@@ -60,7 +60,7 @@ def image_loader(filepath: str):
 
 class ImageDeserializer(Deserializer):
     @requires("image")
-    def deserialize(self, data: str) -> Dict:
+    def serve_load_sample(self, data: str) -> Dict:
         encoded_with_padding = (data + "===").encode("ascii")
         img = base64.b64decode(encoded_with_padding)
         buffer = BytesIO(img)

flash/image/segmentation/data.py

@@ -191,8 +191,8 @@ def load_sample(self, sample: Dict[str, Any]) -> Dict[str, Any]:
 
 
 class SemanticSegmentationDeserializer(ImageDeserializer):
-    def deserialize(self, data: str) -> Dict[str, Any]:
-        result = super().deserialize(data)
+    def serve_load_sample(self, data: str) -> Dict[str, Any]:
+        result = super().serve_load_sample(data)
         result[DataKeys.INPUT] = FT.to_tensor(result[DataKeys.INPUT])
         result[DataKeys.METADATA] = {"size": result[DataKeys.INPUT].shape[-2:]}
         return result

flash/tabular/data.py

@@ -186,7 +186,7 @@ def parameters(self) -> Dict[str, Any]:
             )
         return parameters_state.parameters
 
-    def deserialize(self, data: str) -> Any:
+    def serve_load_sample(self, data: str) -> Any:
         parameters = self.parameters
 
         df = read_csv(StringIO(data))

flash/text/classification/data.py

@@ -47,7 +47,7 @@ def __init__(self, backbone: str, max_length: int, use_fast: bool = True, **kwar
         self.tokenizer = AutoTokenizer.from_pretrained(backbone, use_fast=use_fast, **kwargs)
         self.max_length = max_length
 
-    def deserialize(self, text: str) -> Tensor:
+    def serve_load_sample(self, text: str) -> Tensor:
         return self.tokenizer(text, max_length=self.max_length, truncation=True, padding="max_length")
 
     @property

tests/core/data/io/test_input_base.py

@@ -12,8 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import pytest
+from pytorch_lightning.utilities.exceptions import MisconfigurationException
 
-from flash.core.data.io.input_base import Input, IterableInput
+from flash.core.data.io.input_base import Input, IterableInput, ServeInput
 from flash.core.utilities.stages import RunningStage
 
 
@@ -55,3 +56,28 @@ def __init__(self, *args, **kwargs):
             self.data = iter([1, 2, 3])
 
     ValidIterableInput(RunningStage.TRAINING)
+
+
+def test_serve_input():
+
+    server_input = ServeInput()
+    assert server_input.serving
+    with pytest.raises(NotImplementedError):
+        server_input._call_load_sample("")
+
+    class CustomServeInput(ServeInput):
+        def serve_load_data(self, data):
+            raise NotImplementedError
+
+        def serve_load_sample(self, data):
+            return data + 1
+
+    with pytest.raises(MisconfigurationException, match="serve_load_data"):
+        serve_input = CustomServeInput()
+
+    class CustomServeInput2(ServeInput):
+        def serve_load_sample(self, data):
+            return data + 1
+
+    serve_input = CustomServeInput2()
+    assert serve_input._call_load_sample(1) == 2