onnxtr 0.2.0__py3-none-any.whl → 0.3.1__py3-none-any.whl

onnxtr/models/detection/models/differentiable_binarization.py

@@ -8,7 +8,7 @@ from typing import Any, Dict, Optional
 import numpy as np
 from scipy.special import expit
 
-from ...engine import Engine
+from ...engine import Engine, EngineConfig
 from ..postprocessor.base import GeneralDetectionPostProcessor
 
 __all__ = ["DBNet", "db_resnet50", "db_resnet34", "db_mobilenet_v3_large"]
@@ -33,8 +33,8 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 1024, 1024),
         "mean": (0.798, 0.785, 0.772),
         "std": (0.264, 0.2749, 0.287),
-        "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/db_mobilenet_v3_large-1866973f.onnx",
-        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/db_mobilenet_v3_large_static_8_bit-51659bb9.onnx",
+        "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.2.0/db_mobilenet_v3_large-4987e7bd.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.2.0/db_mobilenet_v3_large_static_8_bit-535a6f25.onnx",
     },
 }
 
@@ -45,6 +45,7 @@ class DBNet(Engine):
     Args:
     ----
         model_path: path or url to onnx model file
+        engine_cfg: configuration for the inference engine
         bin_thresh: threshold for binarization of the output feature map
         box_thresh: minimal objectness score to consider a box
         assume_straight_pages: if True, fit straight bounding boxes only
@@ -54,14 +55,15 @@ class DBNet(Engine):
 
     def __init__(
         self,
-        model_path,
+        model_path: str,
+        engine_cfg: Optional[EngineConfig] = None,
         bin_thresh: float = 0.3,
         box_thresh: float = 0.1,
         assume_straight_pages: bool = True,
         cfg: Optional[Dict[str, Any]] = None,
         **kwargs: Any,
     ) -> None:
-        super().__init__(url=model_path, **kwargs)
+        super().__init__(url=model_path, engine_cfg=engine_cfg, **kwargs)
         self.cfg = cfg
         self.assume_straight_pages = assume_straight_pages
         self.postprocessor = GeneralDetectionPostProcessor(
@@ -91,16 +93,20 @@ def _dbnet(
     arch: str,
     model_path: str,
     load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
     **kwargs: Any,
 ) -> DBNet:
     # Patch the url
     model_path = default_cfgs[arch]["url_8_bit"] if load_in_8_bit and "http" in model_path else model_path
     # Build the model
-    return DBNet(model_path, cfg=default_cfgs[arch], **kwargs)
+    return DBNet(model_path, cfg=default_cfgs[arch], engine_cfg=engine_cfg, **kwargs)
 
 
 def db_resnet34(
-    model_path: str = default_cfgs["db_resnet34"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+    model_path: str = default_cfgs["db_resnet34"]["url"],
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
 ) -> DBNet:
     """DBNet as described in `"Real-time Scene Text Detection with Differentiable Binarization"
     <https://arxiv.org/pdf/1911.08947.pdf>`_, using a ResNet-34 backbone.
@@ -115,17 +121,21 @@ def db_resnet34(
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _dbnet("db_resnet34", model_path, load_in_8_bit, **kwargs)
+    return _dbnet("db_resnet34", model_path, load_in_8_bit, engine_cfg, **kwargs)
 
 
 def db_resnet50(
-    model_path: str = default_cfgs["db_resnet50"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+    model_path: str = default_cfgs["db_resnet50"]["url"],
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
 ) -> DBNet:
     """DBNet as described in `"Real-time Scene Text Detection with Differentiable Binarization"
     <https://arxiv.org/pdf/1911.08947.pdf>`_, using a ResNet-50 backbone.
@@ -140,17 +150,21 @@ def db_resnet50(
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _dbnet("db_resnet50", model_path, load_in_8_bit, **kwargs)
+    return _dbnet("db_resnet50", model_path, load_in_8_bit, engine_cfg, **kwargs)
 
 
 def db_mobilenet_v3_large(
-    model_path: str = default_cfgs["db_mobilenet_v3_large"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+    model_path: str = default_cfgs["db_mobilenet_v3_large"]["url"],
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
 ) -> DBNet:
     """DBNet as described in `"Real-time Scene Text Detection with Differentiable Binarization"
     <https://arxiv.org/pdf/1911.08947.pdf>`_, using a MobileNet V3 Large backbone.
@@ -165,10 +179,11 @@ def db_mobilenet_v3_large(
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _dbnet("db_mobilenet_v3_large", model_path, load_in_8_bit, **kwargs)
+    return _dbnet("db_mobilenet_v3_large", model_path, load_in_8_bit, engine_cfg, **kwargs)

onnxtr/models/detection/models/fast.py

@@ -9,7 +9,7 @@ from typing import Any, Dict, Optional
 import numpy as np
 from scipy.special import expit
 
-from ...engine import Engine
+from ...engine import Engine, EngineConfig
 from ..postprocessor.base import GeneralDetectionPostProcessor
 
 __all__ = ["FAST", "fast_tiny", "fast_small", "fast_base"]
@@ -43,6 +43,7 @@ class FAST(Engine):
     Args:
     ----
         model_path: path or url to onnx model file
+        engine_cfg: configuration for the inference engine
         bin_thresh: threshold for binarization of the output feature map
         box_thresh: minimal objectness score to consider a box
         assume_straight_pages: if True, fit straight bounding boxes only
@@ -53,13 +54,14 @@ class FAST(Engine):
     def __init__(
         self,
         model_path: str,
+        engine_cfg: Optional[EngineConfig] = None,
         bin_thresh: float = 0.1,
         box_thresh: float = 0.1,
         assume_straight_pages: bool = True,
         cfg: Optional[Dict[str, Any]] = None,
         **kwargs: Any,
     ) -> None:
-        super().__init__(url=model_path, **kwargs)
+        super().__init__(url=model_path, engine_cfg=engine_cfg, **kwargs)
         self.cfg = cfg
         self.assume_straight_pages = assume_straight_pages
 
@@ -90,15 +92,21 @@ def _fast(
     arch: str,
     model_path: str,
     load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
     **kwargs: Any,
 ) -> FAST:
     if load_in_8_bit:
         logging.warning("FAST models do not support 8-bit quantization yet. Loading full precision model...")
     # Build the model
-    return FAST(model_path, cfg=default_cfgs[arch], **kwargs)
+    return FAST(model_path, cfg=default_cfgs[arch], engine_cfg=engine_cfg, **kwargs)
 
 
-def fast_tiny(model_path: str = default_cfgs["fast_tiny"]["url"], load_in_8_bit: bool = False, **kwargs: Any) -> FAST:
+def fast_tiny(
+    model_path: str = default_cfgs["fast_tiny"]["url"],
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
+) -> FAST:
     """FAST as described in `"FAST: Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation"
     <https://arxiv.org/pdf/2111.02394.pdf>`_, using a tiny TextNet backbone.
 
@@ -112,16 +120,22 @@ def fast_tiny(model_path: str = default_cfgs["fast_tiny"]["url"], load_in_8_bit:
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _fast("fast_tiny", model_path, load_in_8_bit, **kwargs)
+    return _fast("fast_tiny", model_path, load_in_8_bit, engine_cfg, **kwargs)
 
 
-def fast_small(model_path: str = default_cfgs["fast_small"]["url"], load_in_8_bit: bool = False, **kwargs: Any) -> FAST:
+def fast_small(
+    model_path: str = default_cfgs["fast_small"]["url"],
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
+) -> FAST:
     """FAST as described in `"FAST: Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation"
     <https://arxiv.org/pdf/2111.02394.pdf>`_, using a small TextNet backbone.
 
@@ -135,16 +149,22 @@ def fast_small(model_path: str = default_cfgs["fast_small"]["url"], load_in_8_bi
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _fast("fast_small", model_path, load_in_8_bit, **kwargs)
+    return _fast("fast_small", model_path, load_in_8_bit, engine_cfg, **kwargs)
 
 
-def fast_base(model_path: str = default_cfgs["fast_base"]["url"], load_in_8_bit: bool = False, **kwargs: Any) -> FAST:
+def fast_base(
+    model_path: str = default_cfgs["fast_base"]["url"],
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
+) -> FAST:
     """FAST as described in `"FAST: Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation"
     <https://arxiv.org/pdf/2111.02394.pdf>`_, using a base TextNet backbone.
 
@@ -158,10 +178,11 @@ def fast_base(model_path: str = default_cfgs["fast_base"]["url"], load_in_8_bit:
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _fast("fast_base", model_path, load_in_8_bit, **kwargs)
+    return _fast("fast_base", model_path, load_in_8_bit, engine_cfg, **kwargs)

onnxtr/models/detection/models/linknet.py

@@ -8,7 +8,7 @@ from typing import Any, Dict, Optional
 import numpy as np
 from scipy.special import expit
 
-from ...engine import Engine
+from ...engine import Engine, EngineConfig
 from ..postprocessor.base import GeneralDetectionPostProcessor
 
 __all__ = ["LinkNet", "linknet_resnet18", "linknet_resnet34", "linknet_resnet50"]
@@ -45,6 +45,7 @@ class LinkNet(Engine):
     Args:
     ----
         model_path: path or url to onnx model file
+        engine_cfg: configuration for the inference engine
         bin_thresh: threshold for binarization of the output feature map
         box_thresh: minimal objectness score to consider a box
         assume_straight_pages: if True, fit straight bounding boxes only
@@ -55,13 +56,14 @@ class LinkNet(Engine):
     def __init__(
         self,
         model_path: str,
+        engine_cfg: Optional[EngineConfig] = None,
         bin_thresh: float = 0.1,
         box_thresh: float = 0.1,
         assume_straight_pages: bool = True,
         cfg: Optional[Dict[str, Any]] = None,
         **kwargs: Any,
     ) -> None:
-        super().__init__(url=model_path, **kwargs)
+        super().__init__(url=model_path, engine_cfg=engine_cfg, **kwargs)
         self.cfg = cfg
         self.assume_straight_pages = assume_straight_pages
 
@@ -92,16 +94,20 @@ def _linknet(
     arch: str,
     model_path: str,
     load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
     **kwargs: Any,
 ) -> LinkNet:
     # Patch the url
     model_path = default_cfgs[arch]["url_8_bit"] if load_in_8_bit and "http" in model_path else model_path
     # Build the model
-    return LinkNet(model_path, cfg=default_cfgs[arch], **kwargs)
+    return LinkNet(model_path, cfg=default_cfgs[arch], engine_cfg=engine_cfg, **kwargs)
 
 
 def linknet_resnet18(
-    model_path: str = default_cfgs["linknet_resnet18"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+    model_path: str = default_cfgs["linknet_resnet18"]["url"],
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
 ) -> LinkNet:
     """LinkNet as described in `"LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation"
     <https://arxiv.org/pdf/1707.03718.pdf>`_.
@@ -116,17 +122,21 @@ def linknet_resnet18(
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: keyword arguments of the LinkNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _linknet("linknet_resnet18", model_path, load_in_8_bit, **kwargs)
+    return _linknet("linknet_resnet18", model_path, load_in_8_bit, engine_cfg, **kwargs)
 
 
 def linknet_resnet34(
-    model_path: str = default_cfgs["linknet_resnet34"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+    model_path: str = default_cfgs["linknet_resnet34"]["url"],
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
 ) -> LinkNet:
     """LinkNet as described in `"LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation"
     <https://arxiv.org/pdf/1707.03718.pdf>`_.
@@ -141,17 +151,21 @@ def linknet_resnet34(
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: keyword arguments of the LinkNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _linknet("linknet_resnet34", model_path, load_in_8_bit, **kwargs)
+    return _linknet("linknet_resnet34", model_path, load_in_8_bit, engine_cfg, **kwargs)
 
 
 def linknet_resnet50(
-    model_path: str = default_cfgs["linknet_resnet50"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+    model_path: str = default_cfgs["linknet_resnet50"]["url"],
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
 ) -> LinkNet:
     """LinkNet as described in `"LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation"
     <https://arxiv.org/pdf/1707.03718.pdf>`_.
@@ -166,10 +180,11 @@ def linknet_resnet50(
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: keyword arguments of the LinkNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _linknet("linknet_resnet50", model_path, load_in_8_bit, **kwargs)
+    return _linknet("linknet_resnet50", model_path, load_in_8_bit, engine_cfg, **kwargs)

onnxtr/models/detection/postprocessor/base.py

@@ -109,7 +109,7 @@ class GeneralDetectionPostProcessor(DetectionPostProcessor):
         contours, _ = cv2.findContours(bitmap.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
         for contour in contours:
             # Check whether smallest enclosing bounding box is not too small
-            if np.any(contour[:, 0].max(axis=0) - contour[:, 0].min(axis=0) < 2):
+            if np.any(contour[:, 0].max(axis=0) - contour[:, 0].min(axis=0) < 2):  # type: ignore[index]
                 continue
             # Compute objectness
             if self.assume_straight_pages:
@@ -136,9 +136,10 @@ class GeneralDetectionPostProcessor(DetectionPostProcessor):
                 # compute relative box to get rid of img shape
                 _box[:, 0] /= width
                 _box[:, 1] /= height
-                boxes.append(_box)
+                # Add score to box as (0, score)
+                boxes.append(np.vstack([_box, np.array([0.0, score])]))
 
         if not self.assume_straight_pages:
-            return np.clip(np.asarray(boxes), 0, 1) if len(boxes) > 0 else np.zeros((0, 4, 2), dtype=pred.dtype)
+            return np.clip(np.asarray(boxes), 0, 1) if len(boxes) > 0 else np.zeros((0, 5, 2), dtype=pred.dtype)
         else:
             return np.clip(np.asarray(boxes), 0, 1) if len(boxes) > 0 else np.zeros((0, 5), dtype=pred.dtype)

onnxtr/models/detection/predictor/base.py

@@ -7,6 +7,7 @@ from typing import Any, List, Tuple, Union
 
 import numpy as np
 
+from onnxtr.models.detection._utils import _remove_padding
 from onnxtr.models.preprocessor import PreProcessor
 from onnxtr.utils.repr import NestedObject
 
@@ -38,6 +39,11 @@ class DetectionPredictor(NestedObject):
         return_maps: bool = False,
         **kwargs: Any,
     ) -> Union[List[np.ndarray], Tuple[List[np.ndarray], List[np.ndarray]]]:
+        # Extract parameters from the preprocessor
+        preserve_aspect_ratio = self.pre_processor.resize.preserve_aspect_ratio
+        symmetric_pad = self.pre_processor.resize.symmetric_pad
+        assume_straight_pages = self.model.assume_straight_pages
+
         # Dimension check
         if any(page.ndim != 3 for page in pages):
             raise ValueError("incorrect input shape: all pages are expected to be multi-channel 2D images.")
@@ -47,7 +53,15 @@ class DetectionPredictor(NestedObject):
             self.model(batch, return_preds=True, return_model_output=True, **kwargs) for batch in processed_batches
         ]
 
-        preds = [pred for batch in predicted_batches for pred in batch["preds"]]
+        # Remove padding from loc predictions
+        preds = _remove_padding(
+            pages,
+            [pred[0] for batch in predicted_batches for pred in batch["preds"]],
+            preserve_aspect_ratio=preserve_aspect_ratio,
+            symmetric_pad=symmetric_pad,
+            assume_straight_pages=assume_straight_pages,
+        )
+
         if return_maps:
             seg_maps = [pred for batch in predicted_batches for pred in batch["out_map"]]
             return preds, seg_maps

onnxtr/models/detection/zoo.py

@@ -3,9 +3,10 @@
 # This program is licensed under the Apache License 2.0.
 # See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
 
-from typing import Any
+from typing import Any, Optional
 
 from .. import detection
+from ..engine import EngineConfig
 from ..preprocessor import PreProcessor
 from .predictor import DetectionPredictor
 
@@ -25,13 +26,19 @@ ARCHS = [
 
 
 def _predictor(
-    arch: Any, assume_straight_pages: bool = True, load_in_8_bit: bool = False, **kwargs: Any
+    arch: Any,
+    assume_straight_pages: bool = True,
+    load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
+    **kwargs: Any,
 ) -> DetectionPredictor:
     if isinstance(arch, str):
         if arch not in ARCHS:
             raise ValueError(f"unknown architecture '{arch}'")
 
-        _model = detection.__dict__[arch](assume_straight_pages=assume_straight_pages, load_in_8_bit=load_in_8_bit)
+        _model = detection.__dict__[arch](
+            assume_straight_pages=assume_straight_pages, load_in_8_bit=load_in_8_bit, engine_cfg=engine_cfg
+        )
     else:
         if not isinstance(arch, (detection.DBNet, detection.LinkNet, detection.FAST)):
             raise ValueError(f"unknown architecture: {type(arch)}")
@@ -41,7 +48,7 @@ def _predictor(
 
     kwargs["mean"] = kwargs.get("mean", _model.cfg["mean"])
     kwargs["std"] = kwargs.get("std", _model.cfg["std"])
-    kwargs["batch_size"] = kwargs.get("batch_size", 4)
+    kwargs["batch_size"] = kwargs.get("batch_size", 2)
     predictor = DetectionPredictor(
         PreProcessor(_model.cfg["input_shape"][1:], **kwargs),
         _model,
@@ -53,6 +60,7 @@ def detection_predictor(
     arch: Any = "fast_base",
     assume_straight_pages: bool = True,
     load_in_8_bit: bool = False,
+    engine_cfg: Optional[EngineConfig] = None,
     **kwargs: Any,
 ) -> DetectionPredictor:
     """Text detection architecture.
@@ -68,10 +76,11 @@ def detection_predictor(
         arch: name of the architecture or model itself to use (e.g. 'db_resnet50')
         assume_straight_pages: If True, fit straight boxes to the page
         load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        engine_cfg: configuration for the inference engine
         **kwargs: optional keyword arguments passed to the architecture
 
     Returns:
     -------
         Detection predictor
     """
-    return _predictor(arch, assume_straight_pages, load_in_8_bit, **kwargs)
+    return _predictor(arch, assume_straight_pages, load_in_8_bit, engine_cfg=engine_cfg, **kwargs)

onnxtr/models/engine.py

@@ -3,14 +3,79 @@
 # This program is licensed under the Apache License 2.0.
 # See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
 
-from typing import Any, List, Union
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 import numpy as np
-import onnxruntime
+from onnxruntime import (
+    ExecutionMode,
+    GraphOptimizationLevel,
+    InferenceSession,
+    SessionOptions,
+    get_available_providers,
+    get_device,
+)
 
 from onnxtr.utils.data import download_from_url
 from onnxtr.utils.geometry import shape_translate
 
+__all__ = ["EngineConfig"]
+
+
+class EngineConfig:
+    """Implements a configuration class for the engine of a model
+
+    Args:
+    ----
+        providers: list of providers to use for inference ref.: https://onnxruntime.ai/docs/execution-providers/
+        session_options: configuration for the inference session ref.: https://onnxruntime.ai/docs/api/python/api_summary.html#sessionoptions
+    """
+
+    def __init__(
+        self,
+        providers: Optional[Union[List[Tuple[str, Dict[str, Any]]], List[str]]] = None,
+        session_options: Optional[SessionOptions] = None,
+    ):
+        self._providers = providers or self._init_providers()
+        self._session_options = session_options or self._init_sess_opts()
+
+    def _init_providers(self) -> List[Tuple[str, Dict[str, Any]]]:
+        providers: Any = [("CPUExecutionProvider", {"arena_extend_strategy": "kSameAsRequested"})]
+        available_providers = get_available_providers()
+        if "CUDAExecutionProvider" in available_providers and get_device() == "GPU":  # pragma: no cover
+            providers.insert(
+                0,
+                (
+                    "CUDAExecutionProvider",
+                    {
+                        "device_id": 0,
+                        "arena_extend_strategy": "kNextPowerOfTwo",
+                        "cudnn_conv_algo_search": "DEFAULT",
+                        "do_copy_in_default_stream": True,
+                    },
+                ),
+            )
+        return providers
+
+    def _init_sess_opts(self) -> SessionOptions:
+        session_options = SessionOptions()
+        session_options.enable_cpu_mem_arena = True
+        session_options.execution_mode = ExecutionMode.ORT_SEQUENTIAL
+        session_options.graph_optimization_level = GraphOptimizationLevel.ORT_ENABLE_ALL
+        session_options.intra_op_num_threads = -1
+        session_options.inter_op_num_threads = -1
+        return session_options
+
+    @property
+    def providers(self) -> Union[List[Tuple[str, Dict[str, Any]]], List[str]]:
+        return self._providers
+
+    @property
+    def session_options(self) -> SessionOptions:
+        return self._session_options
+
+    def __repr__(self) -> str:
+        return f"EngineConfig(providers={self.providers}"
+
 
 class Engine:
     """Implements an abstract class for the engine of a model
@@ -18,15 +83,16 @@ class Engine:
     Args:
     ----
         url: the url to use to download a model if needed
-        providers: list of providers to use for inference
+        engine_cfg: the configuration of the engine
         **kwargs: additional arguments to be passed to `download_from_url`
     """
 
-    def __init__(
-        self, url: str, providers: List[str] = ["CPUExecutionProvider", "CUDAExecutionProvider"], **kwargs: Any
-    ) -> None:
+    def __init__(self, url: str, engine_cfg: Optional[EngineConfig] = None, **kwargs: Any) -> None:
+        engine_cfg = engine_cfg if isinstance(engine_cfg, EngineConfig) else EngineConfig()
         archive_path = download_from_url(url, cache_subdir="models", **kwargs) if "http" in url else url
-        self.runtime = onnxruntime.InferenceSession(archive_path, providers=providers)
+        self.session_options = engine_cfg.session_options
+        self.providers = engine_cfg.providers
+        self.runtime = InferenceSession(archive_path, providers=self.providers, sess_options=self.session_options)
         self.runtime_inputs = self.runtime.get_inputs()[0]
         self.tf_exported = int(self.runtime_inputs.shape[-1]) == 3
         self.fixed_batch_size: Union[int, str] = self.runtime_inputs.shape[