[feature/Pipeline] deprecate and migrate existing transformers pipelines

README.md

@@ -139,12 +139,12 @@ deepsparse.benchmark [-h] [-b BATCH_SIZE] [-shapes INPUT_SHAPES]
 ## 👩‍💻 NLP Inference Example
 
 ```python
-from deepsparse.transformers import pipeline
+from deepsparse import Pipeline
 
 # SparseZoo model stub or path to ONNX file
 model_path = "zoo:nlp/question_answering/bert-base/pytorch/huggingface/squad/12layer_pruned80_quant-none-vnni"
 
-qa_pipeline = pipeline(
+qa_pipeline = Pipeline.create(
     task="question-answering",
     model_path=model_path,
 )

src/deepsparse/pipeline.py

@@ -131,7 +131,7 @@ def __init__(
             self._engine_args["scheduler"] = scheduler
 
         self._onnx_file_path = self.setup_onnx_file_path()
-        self._engine = self._initialize_engine()
+        self.engine = self._initialize_engine()
         pass
 
     def __call__(self, *args, **kwargs) -> BaseModel:
@@ -350,13 +350,6 @@ def model_path(self) -> str:
         """
         return self._model_path
 
-    @property
-    def engine(self) -> Union[Engine, ORTEngine]:
-        """
-        :return: engine instance used for model forward pass in pipeline
-        """
-        return self._engine
-
     @property
     def engine_args(self) -> Dict[str, Any]:
         """

src/deepsparse/transformers/__init__.py

@@ -120,4 +120,3 @@ def _check_transformers_install():
 from .helpers import *
 from .loaders import *
 from .pipelines import *
-from .server import *

src/deepsparse/transformers/eval_downstream.py

@@ -58,7 +58,7 @@
 
 from tqdm.auto import tqdm
 
-from deepsparse.transformers import pipeline
+from deepsparse import Pipeline
 
 
 from datasets import load_dataset, load_metric  # isort: skip
@@ -79,14 +79,14 @@ def squad_eval(args):
     squad_metrics = load_metric("squad")
 
     # load QA pipeline
-    question_answer = pipeline(
+    question_answer = Pipeline.create(
         task="question-answering",
         model_path=args.onnx_filepath,
         engine_type=args.engine,
         num_cores=args.num_cores,
-        max_length=args.max_sequence_length,
+        sequence_length=args.max_sequence_length,
     )
-    print(f"Engine info: {question_answer.model}")
+    print(f"Engine info: {question_answer.engine}")
 
     for idx, sample in enumerate(tqdm(squad)):
         pred = question_answer(
@@ -96,7 +96,7 @@ def squad_eval(args):
         )
 
         squad_metrics.add_batch(
-            predictions=[{"prediction_text": pred["answer"], "id": sample["id"]}],
+            predictions=[{"prediction_text": pred.answer, "id": sample["id"]}],
             references=[{"answers": sample["answers"], "id": sample["id"]}],
         )
 
@@ -114,19 +114,19 @@ def mnli_eval(args):
     mnli_metrics = load_metric("glue", "mnli")
 
     # load pipeline
-    text_classify = pipeline(
+    text_classify = Pipeline.create(
         task="text-classification",
         model_path=args.onnx_filepath,
         engine_type=args.engine,
         num_cores=args.num_cores,
-        max_length=args.max_sequence_length,
+        sequence_length=args.max_sequence_length,
     )
-    print(f"Engine info: {text_classify.model}")
+    print(f"Engine info: {text_classify.engine}")
 
     for idx, sample in enumerate(tqdm(mnli_matched)):
         pred = text_classify(sample["premise"], sample["hypothesis"])
         mnli_metrics.add_batch(
-            predictions=[int(pred[0]["label"].split("_")[-1])],
+            predictions=[int(pred.labels[0].split("_")[-1])],
             references=[sample["label"]],
         )
 
@@ -152,20 +152,20 @@ def qqp_eval(args):
     qqp_metrics = load_metric("glue", "qqp")
 
     # load pipeline
-    text_classify = pipeline(
+    text_classify = Pipeline.create(
         task="text-classification",
         model_path=args.onnx_filepath,
         engine_type=args.engine,
         num_cores=args.num_cores,
-        max_length=args.max_sequence_length,
+        sequence_length=args.max_sequence_length,
     )
-    print(f"Engine info: {text_classify.model}")
+    print(f"Engine info: {text_classify.engine}")
 
     for idx, sample in enumerate(tqdm(qqp)):
         pred = text_classify([[sample["question1"], sample["question2"]]])
 
         qqp_metrics.add_batch(
-            predictions=[int(pred[0]["label"].split("_")[-1])],
+            predictions=[int(pred.labels[0].split("_")[-1])],
             references=[sample["label"]],
         )
 
@@ -181,22 +181,22 @@ def sst2_eval(args):
     sst2_metrics = load_metric("glue", "sst2")
 
     # load pipeline
-    text_classify = pipeline(
+    text_classify = Pipeline.create(
         task="text-classification",
         model_path=args.onnx_filepath,
         engine_type=args.engine,
         num_cores=args.num_cores,
-        max_length=args.max_sequence_length,
+        sequence_length=args.max_sequence_length,
     )
-    print(f"Engine info: {text_classify.model}")
+    print(f"Engine info: {text_classify.engine}")
 
     for idx, sample in enumerate(tqdm(sst2)):
         pred = text_classify(
             sample["sentence"],
         )
 
         sst2_metrics.add_batch(
-            predictions=[int(pred[0]["label"].split("_")[-1])],
+            predictions=[int(pred.labels[0].split("_")[-1])],
             references=[sample["label"]],
         )