onnxtr 0.1.1__py3-none-any.whl → 0.2.0__py3-none-any.whl

onnxtr/models/classification/models/mobilenet.py

@@ -24,6 +24,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 256, 256),
         "classes": [0, -90, 180, 90],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/mobilenet_v3_small_crop_orientation-5620cf7e.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/mobilenet_v3_small_crop_orientation_static_8_bit-4cfaa621.onnx",
     },
     "mobilenet_v3_small_page_orientation": {
         "mean": (0.694, 0.695, 0.693),
@@ -31,6 +32,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 512, 512),
         "classes": [0, -90, 180, 90],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/mobilenet_v3_small_page_orientation-d3f76d79.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/mobilenet_v3_small_page_orientation_static_8_bit-3e5ef3dc.onnx",
     },
 }
 
@@ -64,14 +66,19 @@ class MobileNetV3(Engine):
 def _mobilenet_v3(
     arch: str,
     model_path: str,
+    load_in_8_bit: bool = False,
     **kwargs: Any,
 ) -> MobileNetV3:
+    # Patch the url
+    model_path = default_cfgs[arch]["url_8_bit"] if load_in_8_bit and "http" in model_path else model_path
     _cfg = deepcopy(default_cfgs[arch])
     return MobileNetV3(model_path, cfg=_cfg, **kwargs)
 
 
 def mobilenet_v3_small_crop_orientation(
-    model_path: str = default_cfgs["mobilenet_v3_small_crop_orientation"]["url"], **kwargs: Any
+    model_path: str = default_cfgs["mobilenet_v3_small_crop_orientation"]["url"],
+    load_in_8_bit: bool = False,
+    **kwargs: Any,
 ) -> MobileNetV3:
     """MobileNetV3-Small architecture as described in
     `"Searching for MobileNetV3",
@@ -86,17 +93,20 @@ def mobilenet_v3_small_crop_orientation(
     Args:
     ----
         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
         **kwargs: keyword arguments of the MobileNetV3 architecture
 
     Returns:
     -------
         MobileNetV3
     """
-    return _mobilenet_v3("mobilenet_v3_small_crop_orientation", model_path, **kwargs)
+    return _mobilenet_v3("mobilenet_v3_small_crop_orientation", model_path, load_in_8_bit, **kwargs)
 
 
 def mobilenet_v3_small_page_orientation(
-    model_path: str = default_cfgs["mobilenet_v3_small_page_orientation"]["url"], **kwargs: Any
+    model_path: str = default_cfgs["mobilenet_v3_small_page_orientation"]["url"],
+    load_in_8_bit: bool = False,
+    **kwargs: Any,
 ) -> MobileNetV3:
     """MobileNetV3-Small architecture as described in
     `"Searching for MobileNetV3",
@@ -111,10 +121,11 @@ def mobilenet_v3_small_page_orientation(
     Args:
     ----
         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
         **kwargs: keyword arguments of the MobileNetV3 architecture
 
     Returns:
     -------
         MobileNetV3
     """
-    return _mobilenet_v3("mobilenet_v3_small_page_orientation", model_path, **kwargs)
+    return _mobilenet_v3("mobilenet_v3_small_page_orientation", model_path, load_in_8_bit, **kwargs)

onnxtr/models/classification/predictor/base.py

@@ -22,6 +22,7 @@ class OrientationPredictor(NestedObject):
     ----
         pre_processor: transform inputs for easier batched model inference
         model: core classification architecture (backbone + classification head)
+        load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
     """
 
     _children_names: List[str] = ["pre_processor", "model"]

onnxtr/models/classification/zoo.py

@@ -14,24 +14,25 @@ __all__ = ["crop_orientation_predictor", "page_orientation_predictor"]
 ORIENTATION_ARCHS: List[str] = ["mobilenet_v3_small_crop_orientation", "mobilenet_v3_small_page_orientation"]
 
 
-def _orientation_predictor(arch: str, **kwargs: Any) -> OrientationPredictor:
+def _orientation_predictor(arch: str, load_in_8_bit: bool = False, **kwargs: Any) -> OrientationPredictor:
     if arch not in ORIENTATION_ARCHS:
         raise ValueError(f"unknown architecture '{arch}'")
 
     # Load directly classifier from backbone
-    _model = classification.__dict__[arch]()
+    _model = classification.__dict__[arch](load_in_8_bit=load_in_8_bit)
     kwargs["mean"] = kwargs.get("mean", _model.cfg["mean"])
     kwargs["std"] = kwargs.get("std", _model.cfg["std"])
     kwargs["batch_size"] = kwargs.get("batch_size", 128 if "crop" in arch else 4)
     input_shape = _model.cfg["input_shape"][1:]
     predictor = OrientationPredictor(
-        PreProcessor(input_shape, preserve_aspect_ratio=True, symmetric_pad=True, **kwargs), _model
+        PreProcessor(input_shape, preserve_aspect_ratio=True, symmetric_pad=True, **kwargs),
+        _model,
     )
     return predictor
 
 
 def crop_orientation_predictor(
-    arch: Any = "mobilenet_v3_small_crop_orientation", **kwargs: Any
+    arch: Any = "mobilenet_v3_small_crop_orientation", load_in_8_bit: bool = False, **kwargs: Any
 ) -> OrientationPredictor:
     """Crop orientation classification architecture.
 
@@ -44,17 +45,18 @@ def crop_orientation_predictor(
     Args:
     ----
         arch: name of the architecture to use (e.g. 'mobilenet_v3_small_crop_orientation')
+        load_in_8_bit: load the 8-bit quantized version of the model
         **kwargs: keyword arguments to be passed to the OrientationPredictor
 
     Returns:
     -------
         OrientationPredictor
     """
-    return _orientation_predictor(arch, **kwargs)
+    return _orientation_predictor(arch, load_in_8_bit, **kwargs)
 
 
 def page_orientation_predictor(
-    arch: Any = "mobilenet_v3_small_page_orientation", **kwargs: Any
+    arch: Any = "mobilenet_v3_small_page_orientation", load_in_8_bit: bool = False, **kwargs: Any
 ) -> OrientationPredictor:
     """Page orientation classification architecture.
 
@@ -67,10 +69,11 @@ def page_orientation_predictor(
     Args:
     ----
         arch: name of the architecture to use (e.g. 'mobilenet_v3_small_page_orientation')
+        load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
         **kwargs: keyword arguments to be passed to the OrientationPredictor
 
     Returns:
     -------
         OrientationPredictor
     """
-    return _orientation_predictor(arch, **kwargs)
+    return _orientation_predictor(arch, load_in_8_bit, **kwargs)

onnxtr/models/detection/models/differentiable_binarization.py

@@ -20,18 +20,21 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "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_resnet50-69ba0015.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/db_resnet50_static_8_bit-09a6104f.onnx",
     },
     "db_resnet34": {
         "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_resnet34-b4873198.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/db_resnet34_static_8_bit-027e2c7f.onnx",
     },
     "db_mobilenet_v3_large": {
         "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",
     },
 }
 
@@ -87,13 +90,18 @@ class DBNet(Engine):
 def _dbnet(
     arch: str,
     model_path: str,
+    load_in_8_bit: bool = False,
     **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)
 
 
-def db_resnet34(model_path: str = default_cfgs["db_resnet34"]["url"], **kwargs: Any) -> DBNet:
+def db_resnet34(
+    model_path: str = default_cfgs["db_resnet34"]["url"], load_in_8_bit: bool = False, **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.
 
@@ -106,16 +114,19 @@ def db_resnet34(model_path: str = default_cfgs["db_resnet34"]["url"], **kwargs:
     Args:
     ----
         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
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _dbnet("db_resnet34", model_path, **kwargs)
+    return _dbnet("db_resnet34", model_path, load_in_8_bit, **kwargs)
 
 
-def db_resnet50(model_path: str = default_cfgs["db_resnet50"]["url"], **kwargs: Any) -> DBNet:
+def db_resnet50(
+    model_path: str = default_cfgs["db_resnet50"]["url"], load_in_8_bit: bool = False, **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.
 
@@ -128,16 +139,19 @@ def db_resnet50(model_path: str = default_cfgs["db_resnet50"]["url"], **kwargs:
     Args:
     ----
         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
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _dbnet("db_resnet50", model_path, **kwargs)
+    return _dbnet("db_resnet50", model_path, load_in_8_bit, **kwargs)
 
 
-def db_mobilenet_v3_large(model_path: str = default_cfgs["db_mobilenet_v3_large"]["url"], **kwargs: Any) -> DBNet:
+def db_mobilenet_v3_large(
+    model_path: str = default_cfgs["db_mobilenet_v3_large"]["url"], load_in_8_bit: bool = False, **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.
 
@@ -150,10 +164,11 @@ def db_mobilenet_v3_large(model_path: str = default_cfgs["db_mobilenet_v3_large"
     Args:
     ----
         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
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _dbnet("db_mobilenet_v3_large", model_path, **kwargs)
+    return _dbnet("db_mobilenet_v3_large", model_path, load_in_8_bit, **kwargs)

onnxtr/models/detection/models/fast.py

@@ -3,6 +3,7 @@
 # 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.
 
+import logging
 from typing import Any, Dict, Optional
 
 import numpy as np
@@ -88,13 +89,16 @@ class FAST(Engine):
 def _fast(
     arch: str,
     model_path: str,
+    load_in_8_bit: bool = False,
     **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)
 
 
-def fast_tiny(model_path: str = default_cfgs["fast_tiny"]["url"], **kwargs: Any) -> FAST:
+def fast_tiny(model_path: str = default_cfgs["fast_tiny"]["url"], load_in_8_bit: bool = False, **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.
 
@@ -107,16 +111,17 @@ def fast_tiny(model_path: str = default_cfgs["fast_tiny"]["url"], **kwargs: Any)
     Args:
     ----
         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
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _fast("fast_tiny", model_path, **kwargs)
+    return _fast("fast_tiny", model_path, load_in_8_bit, **kwargs)
 
 
-def fast_small(model_path: str = default_cfgs["fast_small"]["url"], **kwargs: Any) -> FAST:
+def fast_small(model_path: str = default_cfgs["fast_small"]["url"], load_in_8_bit: bool = False, **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.
 
@@ -129,16 +134,17 @@ def fast_small(model_path: str = default_cfgs["fast_small"]["url"], **kwargs: An
     Args:
     ----
         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
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _fast("fast_small", model_path, **kwargs)
+    return _fast("fast_small", model_path, load_in_8_bit, **kwargs)
 
 
-def fast_base(model_path: str = default_cfgs["fast_base"]["url"], **kwargs: Any) -> FAST:
+def fast_base(model_path: str = default_cfgs["fast_base"]["url"], load_in_8_bit: bool = False, **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.
 
@@ -151,10 +157,11 @@ def fast_base(model_path: str = default_cfgs["fast_base"]["url"], **kwargs: Any)
     Args:
     ----
         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
         **kwargs: keyword arguments of the DBNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _fast("fast_base", model_path, **kwargs)
+    return _fast("fast_base", model_path, load_in_8_bit, **kwargs)

onnxtr/models/detection/models/linknet.py

@@ -20,18 +20,21 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "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/linknet_resnet18-e0e0b9dc.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/linknet_resnet18_static_8_bit-3b3a37dd.onnx",
     },
     "linknet_resnet34": {
         "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/linknet_resnet34-93e39a39.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/linknet_resnet34_static_8_bit-2824329d.onnx",
     },
     "linknet_resnet50": {
         "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/linknet_resnet50-15d8c4ec.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/linknet_resnet50_static_8_bit-65d6b0b8.onnx",
     },
 }
 
@@ -88,13 +91,18 @@ class LinkNet(Engine):
 def _linknet(
     arch: str,
     model_path: str,
+    load_in_8_bit: bool = False,
     **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)
 
 
-def linknet_resnet18(model_path: str = default_cfgs["linknet_resnet18"]["url"], **kwargs: Any) -> LinkNet:
+def linknet_resnet18(
+    model_path: str = default_cfgs["linknet_resnet18"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+) -> LinkNet:
     """LinkNet as described in `"LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation"
     <https://arxiv.org/pdf/1707.03718.pdf>`_.
 
@@ -107,16 +115,19 @@ def linknet_resnet18(model_path: str = default_cfgs["linknet_resnet18"]["url"],
     Args:
     ----
         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
         **kwargs: keyword arguments of the LinkNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _linknet("linknet_resnet18", model_path, **kwargs)
+    return _linknet("linknet_resnet18", model_path, load_in_8_bit, **kwargs)
 
 
-def linknet_resnet34(model_path: str = default_cfgs["linknet_resnet34"]["url"], **kwargs: Any) -> LinkNet:
+def linknet_resnet34(
+    model_path: str = default_cfgs["linknet_resnet34"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+) -> LinkNet:
     """LinkNet as described in `"LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation"
     <https://arxiv.org/pdf/1707.03718.pdf>`_.
 
@@ -129,16 +140,19 @@ def linknet_resnet34(model_path: str = default_cfgs["linknet_resnet34"]["url"],
     Args:
     ----
         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
         **kwargs: keyword arguments of the LinkNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _linknet("linknet_resnet34", model_path, **kwargs)
+    return _linknet("linknet_resnet34", model_path, load_in_8_bit, **kwargs)
 
 
-def linknet_resnet50(model_path: str = default_cfgs["linknet_resnet50"]["url"], **kwargs: Any) -> LinkNet:
+def linknet_resnet50(
+    model_path: str = default_cfgs["linknet_resnet50"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+) -> LinkNet:
     """LinkNet as described in `"LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation"
     <https://arxiv.org/pdf/1707.03718.pdf>`_.
 
@@ -151,10 +165,11 @@ def linknet_resnet50(model_path: str = default_cfgs["linknet_resnet50"]["url"],
     Args:
     ----
         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
         **kwargs: keyword arguments of the LinkNet architecture
 
     Returns:
     -------
         text detection architecture
     """
-    return _linknet("linknet_resnet50", model_path, **kwargs)
+    return _linknet("linknet_resnet50", model_path, load_in_8_bit, **kwargs)

onnxtr/models/detection/zoo.py

@@ -24,12 +24,14 @@ ARCHS = [
 ]
 
 
-def _predictor(arch: Any, assume_straight_pages: bool = True, **kwargs: Any) -> DetectionPredictor:
+def _predictor(
+    arch: Any, assume_straight_pages: bool = True, load_in_8_bit: bool = False, **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)
+        _model = detection.__dict__[arch](assume_straight_pages=assume_straight_pages, load_in_8_bit=load_in_8_bit)
     else:
         if not isinstance(arch, (detection.DBNet, detection.LinkNet, detection.FAST)):
             raise ValueError(f"unknown architecture: {type(arch)}")
@@ -50,6 +52,7 @@ def _predictor(arch: Any, assume_straight_pages: bool = True, **kwargs: Any) ->
 def detection_predictor(
     arch: Any = "fast_base",
     assume_straight_pages: bool = True,
+    load_in_8_bit: bool = False,
     **kwargs: Any,
 ) -> DetectionPredictor:
     """Text detection architecture.
@@ -64,10 +67,11 @@ 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
         **kwargs: optional keyword arguments passed to the architecture
 
     Returns:
     -------
         Detection predictor
     """
-    return _predictor(arch, assume_straight_pages, **kwargs)
+    return _predictor(arch, assume_straight_pages, load_in_8_bit, **kwargs)

onnxtr/models/engine.py

@@ -43,8 +43,8 @@ class Engine:
             inputs = np.broadcast_to(inputs, (self.fixed_batch_size, *inputs.shape))
             # combine the results
             logits = np.concatenate(
-                [self.runtime.run(self.output_name, {"input": batch})[0] for batch in inputs], axis=0
+                [self.runtime.run(self.output_name, {self.runtime_inputs.name: batch})[0] for batch in inputs], axis=0
             )
         else:
-            logits = self.runtime.run(self.output_name, {"input": inputs})[0]
+            logits = self.runtime.run(self.output_name, {self.runtime_inputs.name: inputs})[0]
         return shape_translate(logits, format="BHWC")

onnxtr/models/predictor/base.py

@@ -31,6 +31,7 @@ class _OCRPredictor:
             accordingly. Doing so will improve performances for documents with page-uniform rotations.
         preserve_aspect_ratio: if True, resize preserving the aspect ratio (with padding)
         symmetric_pad: if True and preserve_aspect_ratio is True, pas the image symmetrically.
+        load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
         **kwargs: keyword args of `DocumentBuilder`
     """
 
@@ -42,11 +43,14 @@ class _OCRPredictor:
         straighten_pages: bool = False,
         preserve_aspect_ratio: bool = True,
         symmetric_pad: bool = True,
+        load_in_8_bit: bool = False,
         **kwargs: Any,
     ) -> None:
         self.assume_straight_pages = assume_straight_pages
         self.straighten_pages = straighten_pages
-        self.crop_orientation_predictor = None if assume_straight_pages else crop_orientation_predictor()
+        self.crop_orientation_predictor = (
+            None if assume_straight_pages else crop_orientation_predictor(load_in_8_bit=load_in_8_bit)
+        )
         self.doc_builder = DocumentBuilder(**kwargs)
         self.preserve_aspect_ratio = preserve_aspect_ratio
         self.symmetric_pad = symmetric_pad

onnxtr/models/recognition/models/crnn.py

@@ -24,6 +24,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 32, 128),
         "vocab": VOCABS["legacy_french"],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/crnn_vgg16_bn-662979cc.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/crnn_vgg16_bn_static_8_bit-bce050c7.onnx",
     },
     "crnn_mobilenet_v3_small": {
         "mean": (0.694, 0.695, 0.693),
@@ -31,6 +32,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 32, 128),
         "vocab": VOCABS["french"],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/crnn_mobilenet_v3_small-bded4d49.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/crnn_mobilenet_v3_small_static_8_bit-4949006f.onnx",
     },
     "crnn_mobilenet_v3_large": {
         "mean": (0.694, 0.695, 0.693),
@@ -38,6 +40,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 32, 128),
         "vocab": VOCABS["french"],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/crnn_mobilenet_v3_large-d42e8185.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/crnn_mobilenet_v3_large_static_8_bit-459e856d.onnx",
     },
 }
 
@@ -148,6 +151,7 @@ class CRNN(Engine):
 def _crnn(
     arch: str,
     model_path: str,
+    load_in_8_bit: bool = False,
     **kwargs: Any,
 ) -> CRNN:
     kwargs["vocab"] = kwargs.get("vocab", default_cfgs[arch]["vocab"])
@@ -155,12 +159,16 @@ def _crnn(
     _cfg = deepcopy(default_cfgs[arch])
     _cfg["vocab"] = kwargs["vocab"]
     _cfg["input_shape"] = kwargs.get("input_shape", default_cfgs[arch]["input_shape"])
+    # 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 CRNN(model_path, cfg=_cfg, **kwargs)
 
 
-def crnn_vgg16_bn(model_path: str = default_cfgs["crnn_vgg16_bn"]["url"], **kwargs: Any) -> CRNN:
+def crnn_vgg16_bn(
+    model_path: str = default_cfgs["crnn_vgg16_bn"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+) -> CRNN:
     """CRNN with a VGG-16 backbone as described in `"An End-to-End Trainable Neural Network for Image-based
     Sequence Recognition and Its Application to Scene Text Recognition" <https://arxiv.org/pdf/1507.05717.pdf>`_.
 
@@ -173,16 +181,19 @@ def crnn_vgg16_bn(model_path: str = default_cfgs["crnn_vgg16_bn"]["url"], **kwar
     Args:
     ----
         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
         **kwargs: keyword arguments of the CRNN architecture
 
     Returns:
     -------
         text recognition architecture
     """
-    return _crnn("crnn_vgg16_bn", model_path, **kwargs)
+    return _crnn("crnn_vgg16_bn", model_path, load_in_8_bit, **kwargs)
 
 
-def crnn_mobilenet_v3_small(model_path: str = default_cfgs["crnn_mobilenet_v3_small"]["url"], **kwargs: Any) -> CRNN:
+def crnn_mobilenet_v3_small(
+    model_path: str = default_cfgs["crnn_mobilenet_v3_small"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+) -> CRNN:
     """CRNN with a MobileNet V3 Small backbone as described in `"An End-to-End Trainable Neural Network for Image-based
     Sequence Recognition and Its Application to Scene Text Recognition" <https://arxiv.org/pdf/1507.05717.pdf>`_.
 
@@ -195,16 +206,19 @@ def crnn_mobilenet_v3_small(model_path: str = default_cfgs["crnn_mobilenet_v3_sm
     Args:
     ----
         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
         **kwargs: keyword arguments of the CRNN architecture
 
     Returns:
     -------
         text recognition architecture
     """
-    return _crnn("crnn_mobilenet_v3_small", model_path, **kwargs)
+    return _crnn("crnn_mobilenet_v3_small", model_path, load_in_8_bit, **kwargs)
 
 
-def crnn_mobilenet_v3_large(model_path: str = default_cfgs["crnn_mobilenet_v3_large"]["url"], **kwargs: Any) -> CRNN:
+def crnn_mobilenet_v3_large(
+    model_path: str = default_cfgs["crnn_mobilenet_v3_large"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+) -> CRNN:
     """CRNN with a MobileNet V3 Large backbone as described in `"An End-to-End Trainable Neural Network for Image-based
     Sequence Recognition and Its Application to Scene Text Recognition" <https://arxiv.org/pdf/1507.05717.pdf>`_.
 
@@ -217,10 +231,11 @@ def crnn_mobilenet_v3_large(model_path: str = default_cfgs["crnn_mobilenet_v3_la
     Args:
     ----
         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
         **kwargs: keyword arguments of the CRNN architecture
 
     Returns:
     -------
         text recognition architecture
     """
-    return _crnn("crnn_mobilenet_v3_large", model_path, **kwargs)
+    return _crnn("crnn_mobilenet_v3_large", model_path, load_in_8_bit, **kwargs)

onnxtr/models/recognition/models/master.py

@@ -24,6 +24,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 32, 128),
         "vocab": VOCABS["french"],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/master-b1287fcd.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/master_dynamic_8_bit-d8bd8206.onnx",
     },
 }
 
@@ -112,6 +113,7 @@ class MASTERPostProcessor(RecognitionPostProcessor):
 def _master(
     arch: str,
     model_path: str,
+    load_in_8_bit: bool = False,
     **kwargs: Any,
 ) -> MASTER:
     # Patch the config
@@ -120,11 +122,13 @@ def _master(
     _cfg["vocab"] = kwargs.get("vocab", _cfg["vocab"])
 
     kwargs["vocab"] = _cfg["vocab"]
+    # Patch the url
+    model_path = default_cfgs[arch]["url_8_bit"] if load_in_8_bit and "http" in model_path else model_path
 
     return MASTER(model_path, cfg=_cfg, **kwargs)
 
 
-def master(model_path: str = default_cfgs["master"]["url"], **kwargs: Any) -> MASTER:
+def master(model_path: str = default_cfgs["master"]["url"], load_in_8_bit: bool = False, **kwargs: Any) -> MASTER:
     """MASTER as described in paper: <https://arxiv.org/pdf/1910.02562.pdf>`_.
 
     >>> import numpy as np
@@ -136,10 +140,11 @@ def master(model_path: str = default_cfgs["master"]["url"], **kwargs: Any) -> MA
     Args:
     ----
         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
         **kwargs: keywoard arguments passed to the MASTER architecture
 
     Returns:
     -------
         text recognition architecture
     """
-    return _master("master", model_path, **kwargs)
+    return _master("master", model_path, load_in_8_bit, **kwargs)

onnxtr/models/recognition/models/parseq.py

@@ -23,6 +23,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 32, 128),
         "vocab": VOCABS["french"],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/parseq-00b40714.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/parseq_dynamic_8_bit-5b04d9f7.onnx",
     },
 }
 
@@ -32,6 +33,7 @@ class PARSeq(Engine):
 
     Args:
     ----
+        model_path: path to onnx model file
         vocab: vocabulary used for encoding
         cfg: dictionary containing information about the model
         **kwargs: additional arguments to be passed to `Engine`
@@ -99,6 +101,7 @@ class PARSeqPostProcessor(RecognitionPostProcessor):
 def _parseq(
     arch: str,
     model_path: str,
+    load_in_8_bit: bool = False,
     **kwargs: Any,
 ) -> PARSeq:
     # Patch the config
@@ -107,12 +110,14 @@ def _parseq(
     _cfg["input_shape"] = kwargs.get("input_shape", _cfg["input_shape"])
 
     kwargs["vocab"] = _cfg["vocab"]
+    # 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 PARSeq(model_path, cfg=_cfg, **kwargs)
 
 
-def parseq(model_path: str = default_cfgs["parseq"]["url"], **kwargs: Any) -> PARSeq:
+def parseq(model_path: str = default_cfgs["parseq"]["url"], load_in_8_bit: bool = False, **kwargs: Any) -> PARSeq:
     """PARSeq architecture from
     `"Scene Text Recognition with Permuted Autoregressive Sequence Models" <https://arxiv.org/pdf/2207.06966>`_.
 
@@ -125,10 +130,11 @@ def parseq(model_path: str = default_cfgs["parseq"]["url"], **kwargs: Any) -> PA
     Args:
     ----
         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
         **kwargs: keyword arguments of the PARSeq architecture
 
     Returns:
     -------
         text recognition architecture
     """
-    return _parseq("parseq", model_path, **kwargs)
+    return _parseq("parseq", model_path, load_in_8_bit, **kwargs)

onnxtr/models/recognition/models/sar.py

@@ -23,6 +23,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 32, 128),
         "vocab": VOCABS["french"],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/sar_resnet31-395f8005.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/sar_resnet31_static_8_bit-c07316bc.onnx",
     },
 }
 
@@ -99,6 +100,7 @@ class SARPostProcessor(RecognitionPostProcessor):
 def _sar(
     arch: str,
     model_path: str,
+    load_in_8_bit: bool = False,
     **kwargs: Any,
 ) -> SAR:
     # Patch the config
@@ -107,12 +109,16 @@ def _sar(
     _cfg["input_shape"] = kwargs.get("input_shape", _cfg["input_shape"])
 
     kwargs["vocab"] = _cfg["vocab"]
+    # 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 SAR(model_path, cfg=_cfg, **kwargs)
 
 
-def sar_resnet31(model_path: str = default_cfgs["sar_resnet31"]["url"], **kwargs: Any) -> SAR:
+def sar_resnet31(
+    model_path: str = default_cfgs["sar_resnet31"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+) -> SAR:
     """SAR with a resnet-31 feature extractor as described in `"Show, Attend and Read:A Simple and Strong
     Baseline for Irregular Text Recognition" <https://arxiv.org/pdf/1811.00751.pdf>`_.
 
@@ -125,10 +131,11 @@ def sar_resnet31(model_path: str = default_cfgs["sar_resnet31"]["url"], **kwargs
     Args:
     ----
         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
         **kwargs: keyword arguments of the SAR architecture
 
     Returns:
     -------
         text recognition architecture
     """
-    return _sar("sar_resnet31", model_path, **kwargs)
+    return _sar("sar_resnet31", model_path, load_in_8_bit, **kwargs)

onnxtr/models/recognition/models/vitstr.py

@@ -23,6 +23,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 32, 128),
         "vocab": VOCABS["french"],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/vitstr_small-3ff9c500.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/vitstr_small_dynamic_8_bit-bec6c796.onnx",
     },
     "vitstr_base": {
         "mean": (0.694, 0.695, 0.693),
@@ -30,6 +31,7 @@ default_cfgs: Dict[str, Dict[str, Any]] = {
         "input_shape": (3, 32, 128),
         "vocab": VOCABS["french"],
         "url": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.0.1/vitstr_base-ff62f5be.onnx",
+        "url_8_bit": "https://github.com/felixdittrich92/OnnxTR/releases/download/v0.1.2/vitstr_base_dynamic_8_bit-976c7cd6.onnx",
     },
 }
 
@@ -109,6 +111,7 @@ class ViTSTRPostProcessor(RecognitionPostProcessor):
 def _vitstr(
     arch: str,
     model_path: str,
+    load_in_8_bit: bool = False,
     **kwargs: Any,
 ) -> ViTSTR:
     # Patch the config
@@ -117,12 +120,16 @@ def _vitstr(
     _cfg["input_shape"] = kwargs.get("input_shape", _cfg["input_shape"])
 
     kwargs["vocab"] = _cfg["vocab"]
+    # 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 ViTSTR(model_path, cfg=_cfg, **kwargs)
 
 
-def vitstr_small(model_path: str = default_cfgs["vitstr_small"]["url"], **kwargs: Any) -> ViTSTR:
+def vitstr_small(
+    model_path: str = default_cfgs["vitstr_small"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+) -> ViTSTR:
     """ViTSTR-Small as described in `"Vision Transformer for Fast and Efficient Scene Text Recognition"
     <https://arxiv.org/pdf/2105.08582.pdf>`_.
 
@@ -135,16 +142,19 @@ def vitstr_small(model_path: str = default_cfgs["vitstr_small"]["url"], **kwargs
     Args:
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
-        kwargs: keyword arguments of the ViTSTR architecture
+        load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        **kwargs: keyword arguments of the ViTSTR architecture
 
     Returns:
     -------
         text recognition architecture
     """
-    return _vitstr("vitstr_small", model_path, **kwargs)
+    return _vitstr("vitstr_small", model_path, load_in_8_bit, **kwargs)
 
 
-def vitstr_base(model_path: str = default_cfgs["vitstr_base"]["url"], **kwargs: Any) -> ViTSTR:
+def vitstr_base(
+    model_path: str = default_cfgs["vitstr_base"]["url"], load_in_8_bit: bool = False, **kwargs: Any
+) -> ViTSTR:
     """ViTSTR-Base as described in `"Vision Transformer for Fast and Efficient Scene Text Recognition"
     <https://arxiv.org/pdf/2105.08582.pdf>`_.
 
@@ -157,10 +167,11 @@ def vitstr_base(model_path: str = default_cfgs["vitstr_base"]["url"], **kwargs:
     Args:
     ----
         model_path: path to onnx model file, defaults to url in default_cfgs
-        kwargs: keyword arguments of the ViTSTR architecture
+        load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
+        **kwargs: keyword arguments of the ViTSTR architecture
 
     Returns:
     -------
         text recognition architecture
     """
-    return _vitstr("vitstr_base", model_path, **kwargs)
+    return _vitstr("vitstr_base", model_path, load_in_8_bit, **kwargs)

onnxtr/models/recognition/zoo.py

@@ -25,12 +25,12 @@ ARCHS: List[str] = [
 ]
 
 
-def _predictor(arch: Any, **kwargs: Any) -> RecognitionPredictor:
+def _predictor(arch: Any, load_in_8_bit: bool = False, **kwargs: Any) -> RecognitionPredictor:
     if isinstance(arch, str):
         if arch not in ARCHS:
             raise ValueError(f"unknown architecture '{arch}'")
 
-        _model = recognition.__dict__[arch]()
+        _model = recognition.__dict__[arch](load_in_8_bit=load_in_8_bit)
     else:
         if not isinstance(
             arch, (recognition.CRNN, recognition.SAR, recognition.MASTER, recognition.ViTSTR, recognition.PARSeq)
@@ -47,7 +47,9 @@ def _predictor(arch: Any, **kwargs: Any) -> RecognitionPredictor:
     return predictor
 
 
-def recognition_predictor(arch: Any = "crnn_vgg16_bn", **kwargs: Any) -> RecognitionPredictor:
+def recognition_predictor(
+    arch: Any = "crnn_vgg16_bn", load_in_8_bit: bool = False, **kwargs: Any
+) -> RecognitionPredictor:
     """Text recognition architecture.
 
     Example::
@@ -60,10 +62,11 @@ def recognition_predictor(arch: Any = "crnn_vgg16_bn", **kwargs: Any) -> Recogni
     Args:
     ----
         arch: name of the architecture or model itself to use (e.g. 'crnn_vgg16_bn')
+        load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
         **kwargs: optional parameters to be passed to the architecture
 
     Returns:
     -------
         Recognition predictor
     """
-    return _predictor(arch, **kwargs)
+    return _predictor(arch, load_in_8_bit, **kwargs)

onnxtr/models/zoo.py

@@ -23,6 +23,7 @@ def _predictor(
     detect_orientation: bool = False,
     straighten_pages: bool = False,
     detect_language: bool = False,
+    load_in_8_bit: bool = False,
     **kwargs,
 ) -> OCRPredictor:
     # Detection
@@ -32,12 +33,14 @@ def _predictor(
         assume_straight_pages=assume_straight_pages,
         preserve_aspect_ratio=preserve_aspect_ratio,
         symmetric_pad=symmetric_pad,
+        load_in_8_bit=load_in_8_bit,
     )
 
     # Recognition
     reco_predictor = recognition_predictor(
         reco_arch,
         batch_size=reco_bs,
+        load_in_8_bit=load_in_8_bit,
     )
 
     return OCRPredictor(
@@ -63,6 +66,7 @@ def ocr_predictor(
     detect_orientation: bool = False,
     straighten_pages: bool = False,
     detect_language: bool = False,
+    load_in_8_bit: bool = False,
     **kwargs: Any,
 ) -> OCRPredictor:
     """End-to-end OCR architecture using one model for localization, and another for text recognition.
@@ -94,6 +98,7 @@ def ocr_predictor(
             Doing so will improve performances for documents with page-uniform rotations.
         detect_language: if True, the language prediction will be added to the predictions for each
             page. Doing so will slightly deteriorate the overall latency.
+        load_in_8_bit: whether to load the the 8-bit quantized model, defaults to False
         kwargs: keyword args of `OCRPredictor`
 
     Returns:
@@ -110,5 +115,6 @@ def ocr_predictor(
         detect_orientation=detect_orientation,
         straighten_pages=straighten_pages,
         detect_language=detect_language,
+        load_in_8_bit=load_in_8_bit,
         **kwargs,
     )

onnxtr/version.py

@@ -1 +1 @@
-__version__ = 'v0.1.1'
+__version__ = 'v0.2.0'

{onnxtr-0.1.1.dist-info → onnxtr-0.2.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: onnxtr
-Version: 0.1.1
+Version: 0.2.0
 Summary: Onnx Text Recognition (OnnxTR): docTR Onnx-Wrapper for high-performance OCR on documents.
 Author-email: Felix Dittrich <felixdittrich92@gmail.com>
 Maintainer: Felix Dittrich
@@ -227,7 +227,6 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: numpy <2.0.0,>=1.16.0
 Requires-Dist: scipy <2.0.0,>=1.4.0
-Requires-Dist: onnx <2.0.0,>=1.12.0
 Requires-Dist: opencv-python <5.0.0,>=4.5.0
 Requires-Dist: pypdfium2 <5.0.0,>=4.0.0
 Requires-Dist: pyclipper <2.0.0,>=1.2.0
@@ -276,17 +275,18 @@ Requires-Dist: mplcursors >=0.3 ; extra == 'viz'
 [![codecov](https://codecov.io/gh/felixdittrich92/OnnxTR/graph/badge.svg?token=WVFRCQBOLI)](https://codecov.io/gh/felixdittrich92/OnnxTR)
 [![Codacy Badge](https://app.codacy.com/project/badge/Grade/4fff4d764bb14fb8b4f4afeb9587231b)](https://app.codacy.com/gh/felixdittrich92/OnnxTR/dashboard?utm_source=gh&utm_medium=referral&utm_content=&utm_campaign=Badge_grade)
 [![CodeFactor](https://www.codefactor.io/repository/github/felixdittrich92/onnxtr/badge)](https://www.codefactor.io/repository/github/felixdittrich92/onnxtr)
-[![Pypi](https://img.shields.io/badge/pypi-v0.0.1-blue.svg)](https://pypi.org/project/OnnxTR/)
+[![Pypi](https://img.shields.io/badge/pypi-v0.2.0-blue.svg)](https://pypi.org/project/OnnxTR/)
 
-> :warning: Please note that this is wrapper around the [doctr](https://github.com/mindee/doctr) library to provide a Onnx pipeline for docTR. For feature requests, which are not directly related to the Onnx pipeline, please refer to the base project.
+> :warning: Please note that this is a wrapper around the [doctr](https://github.com/mindee/doctr) library to provide a Onnx pipeline for docTR. For feature requests, which are not directly related to the Onnx pipeline, please refer to the base project.
 
 **Optical Character Recognition made seamless & accessible to anyone, powered by Onnx**
 
 What you can expect from this repository:
 
 - efficient ways to parse textual information (localize and identify each word) from your documents
-- a Onnx pipeline for docTR, a wrapper around the [doctr](https://github.com/mindee/doctr) library
+- a Onnx pipeline for docTR, a wrapper around the [doctr](https://github.com/mindee/doctr) library - no PyTorch or TensorFlow dependencies
 - more lightweight package with faster inference latency and less required resources
+- 8-Bit quantized models for faster inference on CPU
 
 ![OCR_example](https://github.com/felixdittrich92/OnnxTR/raw/main/docs/images/ocr.png)
 
@@ -300,10 +300,14 @@ Python 3.9 (or higher) and [pip](https://pip.pypa.io/en/stable/) are required to
 
 You can then install the latest release of the package using [pypi](https://pypi.org/project/OnnxTR/) as follows:
 
-**NOTE:** For GPU support please take a look at: [ONNX Runtime](https://onnxruntime.ai/getting-started). Currently supported execution providers by default are: CPU, CUDA
+**NOTE:**
+
+For GPU support please take a look at: [ONNX Runtime](https://onnxruntime.ai/getting-started). Currently supported execution providers by default are: CPU, CUDA
+
+- **Prerequisites:** CUDA & cuDNN needs to be installed before [Version table](https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html).
 
 ```shell
-pip install onnxtr[cpu]
+pip install "onnxtr[cpu]"
 # with gpu support
 pip install "onnxtr[gpu]"
 # with HTML support
@@ -340,17 +344,24 @@ from onnxtr.models import ocr_predictor
 
 model = ocr_predictor(
     det_arch='fast_base',  # detection architecture
-    rec_arch='vitstr_base',  # recognition architecture
+    reco_arch='vitstr_base',  # recognition architecture
     det_bs=4, # detection batch size
     reco_bs=1024, # recognition batch size
     assume_straight_pages=True,  # set to `False` if the pages are not straight (rotation, perspective, etc.) (default: True)
     straighten_pages=False,  # set to `True` if the pages should be straightened before final processing (default: False)
+    # Preprocessing related parameters
     preserve_aspect_ratio=True,  # set to `False` if the aspect ratio should not be preserved (default: True)
     symmetric_pad=True,  # set to `False` to disable symmetric padding (default: True)
+    # Additional parameters - meta information
+    detect_orientation=False,  # set to `True` if the orientation of the pages should be detected (default: False)
+    detect_language=False, # set to `True` if the language of the pages should be detected (default: False)
     # DocumentBuilder specific parameters
     resolve_lines=True,  # whether words should be automatically grouped into lines (default: True)
     resolve_blocks=True,  # whether lines should be automatically grouped into blocks (default: True)
     paragraph_break=0.035,  # relative length of the minimum space separating paragraphs (default: 0.035)
+    # OnnxTR specific parameters
+    # NOTE: 8-Bit quantized models are not available for FAST detection models and can in general lead to poorer accuracy
+    load_in_8_bit=False,  # set to `True` to load 8-bit quantized models instead of the full precision onces (default: False)
 )
 # PDF
 doc = DocumentFile.from_pdf("path/to/your/doc.pdf")
@@ -398,7 +409,7 @@ from onnxtr.models import ocr_predictor, linknet_resnet18, parseq
 
 reco_model = parseq("path_to_custom_model.onnx", vocab="ABC")
 det_model = linknet_resnet18("path_to_custom_model.onnx")
-model = ocr_predictor(det_model=det_model, reco_model=reco_model)
+model = ocr_predictor(det_arch=det_model, reco_arch=reco_model)
 ```
 
 ## Models architectures
@@ -431,9 +442,9 @@ predictor.list_archs()
             'linknet_resnet18',
             'linknet_resnet34',
             'linknet_resnet50',
-            'fast_tiny',
-            'fast_small',
-            'fast_base'
+            'fast_tiny',  # No 8-bit support
+            'fast_small',  # No 8-bit support
+            'fast_base'  # No 8-bit support
         ],
     'recognition archs':
         [
@@ -462,7 +473,36 @@ NOTE:
 
 ### Benchmarks
 
-COMING SOON
+The CPU benchmarks was measured on a `i7-14700K Intel CPU`.
+
+The GPU benchmarks was measured on a `RTX 4080 Nvidia GPU`.
+
+Benchmarking performed on the FUNSD dataset and CORD dataset.
+
+docTR / OnnxTR models used for the benchmarks are `fast_base` (full precision) | `db_resnet50` (8-bit variant) for detection and `crnn_vgg16_bn` for recognition.
+
+The smallest combination in OnnxTR (docTR) of `db_mobilenet_v3_large` and `crnn_mobilenet_v3_small` takes as comparison `~0.17s / Page` on the FUNSD dataset and `~0.12s / Page` on the CORD dataset in **full precision**.
+
+- CPU benchmarks:
+
+|Library                         |FUNSD (199 pages)              |CORD  (900 pages)              |
+|--------------------------------|-------------------------------|-------------------------------|
+|docTR (CPU) - v0.8.1            | ~1.29s / Page                 | ~0.60s / Page                 |
+|**OnnxTR (CPU)** - v0.1.2       | ~0.57s / Page                 | **~0.25s / Page**             |
+|**OnnxTR (CPU) 8-bit** - v0.1.2 | **~0.38s / Page**             | **~0.14s / Page**             |
+|EasyOCR (CPU) - v1.7.1          | ~1.96s / Page                 | ~1.75s / Page                 |
+|**PyTesseract (CPU)** - v0.3.10 | **~0.50s / Page**             | ~0.52s / Page                 |
+|Surya (line) (CPU) - v0.4.4     | ~48.76s / Page                | ~35.49s / Page                |
+
+- GPU benchmarks:
+
+|Library                         |FUNSD (199 pages)              |CORD  (900 pages)              |
+|--------------------------------|-------------------------------|-------------------------------|
+|docTR (GPU) - v0.8.1            | ~0.07s / Page                 | ~0.05s / Page                 |
+|**docTR (GPU) float16** - v0.8.1| **~0.06s / Page**             | **~0.03s / Page**             |
+|OnnxTR (GPU) - v0.1.2           | **~0.06s / Page**             | ~0.04s / Page                 |
+|EasyOCR (GPU) - v1.7.1          | ~0.31s / Page                 | ~0.19s / Page                 |
+|Surya (GPU) float16 - v0.4.4    | ~3.70s / Page                 | ~2.81s / Page                 |
 
 ## Citation
 

{onnxtr-0.1.1.dist-info → onnxtr-0.2.0.dist-info}/RECORD

@@ -1,6 +1,6 @@
 onnxtr/__init__.py,sha256=h7Wc2tuHLsaoCk5xNpEFEK-g11A6SJA7nAasA76TQ_Y,100
 onnxtr/file_utils.py,sha256=WjUKalEdR53aoeIY4e-ihy3r7J_C9qFxL40JHGPfutc,1107
-onnxtr/version.py,sha256=jQY6B0xUe16A1ACYL5noxMrvy3wwP2XMsFq2Su0alUA,23
+onnxtr/version.py,sha256=F-dLDdnsmwBrwqQCEb3nX0LORU_q1xg9pmFGVU_3j8o,23
 onnxtr/contrib/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnxtr/contrib/artefacts.py,sha256=tdmfhvfXVRYEH7uj4_hqf2cuUGoTieyNK8bXsD3zHwo,5383
 onnxtr/contrib/base.py,sha256=PoCKtOIgj7u4xl-V0eBVh-QmVeTyk_eEggFHQ8R34AI,3445
@@ -13,40 +13,40 @@ onnxtr/io/reader.py,sha256=BA7DPhW-Gkmce_ZfzrOl4H3pSXVy2JBeQEuY3pWrBFg,2852
 onnxtr/models/__init__.py,sha256=CzdiA34zjFq4a22XvgzG_Bojosi9aN5jeRCizRyh_7o,101
 onnxtr/models/_utils.py,sha256=RnSmj70gFU9CCJlsCYiVULEs7L8MhogkLf2Xu_7wBhM,4824
 onnxtr/models/builder.py,sha256=byUpCs9x5RS2lqNXLECR5GELqd1yF-MEg_09OGWj20Q,13642
-onnxtr/models/engine.py,sha256=2SwzcWCf2s4ti1mHqLF0lqLDw3wpM9FYjdEdtK5MYJ4,2140
-onnxtr/models/zoo.py,sha256=FTG6KCv-4G90N-m0fZhRDOFiyd_sFzLZuBw4bg2ecvI,4302
+onnxtr/models/engine.py,sha256=EtBIceuxIdJ1bqEl59n35D-AUrkIEelSOqD1IWEtXAE,2174
+onnxtr/models/zoo.py,sha256=oHk9BOt5U3fZBpeFRgKkQ_U6PthwfoKw0lBEi4Lu6ak,4567
 onnxtr/models/classification/__init__.py,sha256=h1bZs55iLJBMATtzS4ntTKwfD6OGXBiiqGv_hEnOFnE,41
-onnxtr/models/classification/zoo.py,sha256=yiDTEZCHLBu6WUwQVnQipKhjngPs6OgTeNt4MX8z9m4,2748
+onnxtr/models/classification/zoo.py,sha256=MW2UH3MOEJ1aDwHQ_SjrUfPb_U6MrWA3lOGlhwxogaQ,3058
 onnxtr/models/classification/models/__init__.py,sha256=rohbM6ZQslfYchi7feZwwh-sX3XXRUhgtEJQeurAytQ,24
-onnxtr/models/classification/models/mobilenet.py,sha256=9SzYYpn1XOUFHZHzKM0BtvSh08WXeAhMeebt3E1eAmQ,3677
+onnxtr/models/classification/models/mobilenet.py,sha256=z9zN0GnXcOCTH-Hu-xKsjA4W7r3D8F9n_hE-T_Rzqhc,4427
 onnxtr/models/classification/predictor/__init__.py,sha256=ERmmOxz_9mUkIuccNbzUa5Y6gVLLVDdyc4cCxbCCUbY,20
-onnxtr/models/classification/predictor/base.py,sha256=9-5i1Uj0EdNQBeSOhaQjGz78Gr0IAr3JZhoKVMA_il4,2009
+onnxtr/models/classification/predictor/base.py,sha256=Xfaj2XlaJuQ2R81XqF5RB0Wcvzd4wh7Z6j1ifn2niFc,2097
 onnxtr/models/detection/__init__.py,sha256=h1bZs55iLJBMATtzS4ntTKwfD6OGXBiiqGv_hEnOFnE,41
 onnxtr/models/detection/core.py,sha256=ZmVDHLJ1l4LQ8rFSKc7enXDkGcOWrcQv4H0SJWyLsag,3584
-onnxtr/models/detection/zoo.py,sha256=GfwvVAT3bbx-1zccqDO-jzOzzL5Sj3D06zbd_x6dUOo,2257
+onnxtr/models/detection/zoo.py,sha256=4DDHWQtWAQpEhl0SHcqEV12HgjXpTBh4LB4tdQc7etk,2457
 onnxtr/models/detection/models/__init__.py,sha256=6Ea6knYrVCR2jAmPlsVWmCdHe-c6lSRETSAuZGfhx8I,85
-onnxtr/models/detection/models/differentiable_binarization.py,sha256=fu1Mn8i38vIyW98YKdPnwXRcvk_-OlkOmcez7wef8l4,5157
-onnxtr/models/detection/models/fast.py,sha256=40P_q5TWUnY9iIyi-pAqnvauQ5KeqIAco2NGGF4plrg,5101
-onnxtr/models/detection/models/linknet.py,sha256=e9uZZB5S3GYImGroHyZ4oonb_ULxL0VAXd_yDuqG1K0,5194
+onnxtr/models/detection/models/differentiable_binarization.py,sha256=mn1kE7k6VLsiyERH9ghtVJBqalG_In2dv8SvtJNN_HM,6131
+onnxtr/models/detection/models/fast.py,sha256=jhw8M6vf9FM8Q5a5Lgu6mfpzkLrIEErX-1GxCR_DfYw,5677
+onnxtr/models/detection/models/linknet.py,sha256=PCwaYhTqFgt2BbixoVTeZWHN9fd4r2i1GTZ0CywZRvw,6173
 onnxtr/models/detection/postprocessor/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 onnxtr/models/detection/postprocessor/base.py,sha256=U8FgqveZt2qPHI0aUnIEeX370zGUEWUxL6EPqhFZnRI,5714
 onnxtr/models/detection/predictor/__init__.py,sha256=ERmmOxz_9mUkIuccNbzUa5Y6gVLLVDdyc4cCxbCCUbY,20
 onnxtr/models/detection/predictor/base.py,sha256=K0fSTFnA8-fGL9F3HzW3nqnEqO5YKbhFf2TvpUOe_Tw,1744
 onnxtr/models/predictor/__init__.py,sha256=XL25XkRkgyK7mldF-CWhg2MMakSdP5vLpDLwL59hphk,25
-onnxtr/models/predictor/base.py,sha256=nrIE6rvxA3LbeUbQ-e3wkochO66OJT3BP-NW_OyHUyw,7555
+onnxtr/models/predictor/base.py,sha256=6rWf8FZrp6ku4yvP1jJl39t0Y8L7wqWb2GYYagTExtw,7731
 onnxtr/models/predictor/predictor.py,sha256=sfwqRzabqavlI6ty_Z6jGm64olmmQBLQGU39_nulQZs,6045
 onnxtr/models/preprocessor/__init__.py,sha256=ERmmOxz_9mUkIuccNbzUa5Y6gVLLVDdyc4cCxbCCUbY,20
 onnxtr/models/preprocessor/base.py,sha256=f0t0rMCzvuxwgq7jlKvcVWyjeDOx7yCLUw52quEaETM,3990
 onnxtr/models/recognition/__init__.py,sha256=h1bZs55iLJBMATtzS4ntTKwfD6OGXBiiqGv_hEnOFnE,41
 onnxtr/models/recognition/core.py,sha256=0Q1dVXqRcDUr_ycT5tpoSH9-zuDF58GtnmxWpUS8Ibo,739
 onnxtr/models/recognition/utils.py,sha256=04abbjx-_OuF5iEANWIAOK3tQQl1tExPmBQx4IG04Lc,3569
-onnxtr/models/recognition/zoo.py,sha256=FC8kHSr66s5D-Ltb8dnNI-x3HLdy5LZhZeiWMV69J9k,2115
+onnxtr/models/recognition/zoo.py,sha256=MR5O5UVM0tmWGKdGzXTE7HTBCYKZBuOVwvRBASz-zRo,2309
 onnxtr/models/recognition/models/__init__.py,sha256=IXfiuzzkft8O1CpBZWYTpFw19y49mt5rJ_iGSdaWiU0,105
-onnxtr/models/recognition/models/crnn.py,sha256=Y1lmvCAN_r-ciZNHRIZWHLhv60j2zcR1XoRC5r8_BD0,7375
-onnxtr/models/recognition/models/master.py,sha256=4FiWWBtezb7WjhDvnzauZ-iL6_rKashoqbs-9JfHlvM,4009
-onnxtr/models/recognition/models/parseq.py,sha256=qYtXWC8jXnIKViKTLp-WvcEOoU_DPkZ-_KHVhoKgUmw,3765
-onnxtr/models/recognition/models/sar.py,sha256=c0hRZ5dXl9sijkXZXtUnu9cB4DjUFc1MMzriDBhor78,3815
-onnxtr/models/recognition/models/vitstr.py,sha256=jID2RJs5O81b4jBGjkeSNPQ2PTbon-_U8ZDGfCoGcwY,4880
+onnxtr/models/recognition/models/crnn.py,sha256=jtrbe2WqPCvWCgaZkWVCX0LPE3poQXQpe6Aq6pJN6i4,8365
+onnxtr/models/recognition/models/master.py,sha256=iacF5zZNf3kWTetYbGemBMey8o0heHmgFi8sVoTMyik,4427
+onnxtr/models/recognition/models/parseq.py,sha256=omHQVfJ1BR03JxiPcr5wxR_cx9okiWddTCikpUOJc-E,4227
+onnxtr/models/recognition/models/sar.py,sha256=12U2RH8jQp4SE2axHvME3TjIa5WkSbZaJfzClFxVGUE,4244
+onnxtr/models/recognition/models/vitstr.py,sha256=uPyBBE-PDPJNxzUFqJbaVdm6MTCbE5dgvst2MSPqswk,5583
 onnxtr/models/recognition/predictor/__init__.py,sha256=ERmmOxz_9mUkIuccNbzUa5Y6gVLLVDdyc4cCxbCCUbY,20
 onnxtr/models/recognition/predictor/_utils.py,sha256=ZNm5I7ibiWfTlz302uiifCkUOu65YWa-oUBUMPrrUuQ,3406
 onnxtr/models/recognition/predictor/base.py,sha256=YvqSNEM3rCEttxl6hsC9zl1R97N9zO2WZfD5_-nfkR0,2483
@@ -62,9 +62,9 @@ onnxtr/utils/reconstitution.py,sha256=Hx1_ddLevKLzuxXc19UelPdsGlAwqi4f6vRSYKHDUB
 onnxtr/utils/repr.py,sha256=kfbjGL6KymGT8spo2UL4FJXZ0XRwa7CO7Y1dTVR8dIk,2129
 onnxtr/utils/visualization.py,sha256=CX09qvDnNIw3BFW5F3jM4R9OcpLWAeZyoDyTAOGRvls,9925
 onnxtr/utils/vocabs.py,sha256=SCQ4XQjbHSxunj1tg2iHRiPfE8OaTAMhcJbKq5BNvFs,3138
-onnxtr-0.1.1.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-onnxtr-0.1.1.dist-info/METADATA,sha256=Fo9T9dyuaLLYzE8cAxCdiIQfhQRnumUqw2X1OS_PqIc,24740
-onnxtr-0.1.1.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-onnxtr-0.1.1.dist-info/top_level.txt,sha256=48aR9EH095hmgye7CNuxAW3o_cj4TjhWmuw3jEMdTs0,12
-onnxtr-0.1.1.dist-info/zip-safe,sha256=AbpHGcgLb-kRsJGnwFEktk7uzpZOCcBY74-YBdrKVGs,1
-onnxtr-0.1.1.dist-info/RECORD,,
+onnxtr-0.2.0.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+onnxtr-0.2.0.dist-info/METADATA,sha256=Uc9NddF-vWik-X2Sohhx-3Jw8gZJxp86GFpt1Q7-QRg,27690
+onnxtr-0.2.0.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+onnxtr-0.2.0.dist-info/top_level.txt,sha256=48aR9EH095hmgye7CNuxAW3o_cj4TjhWmuw3jEMdTs0,12
+onnxtr-0.2.0.dist-info/zip-safe,sha256=AbpHGcgLb-kRsJGnwFEktk7uzpZOCcBY74-YBdrKVGs,1
+onnxtr-0.2.0.dist-info/RECORD,,