paddlex 3.0.0b2__py3-none-any.whl → 3.0.0rc0__py3-none-any.whl

paddlex/inference/models/image_unwarping/predictor.py

@@ -0,0 +1,105 @@
+# copyright (c) 2024 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Union, Dict, List, Tuple
+import numpy as np
+
+from ....modules.image_unwarping.model_list import MODELS
+from ...common.batch_sampler import ImageBatchSampler
+from ...common.reader import ReadImage
+from ..common import (
+    Normalize,
+    ToCHWImage,
+    ToBatch,
+    StaticInfer,
+)
+from ..base import BasicPredictor
+from .processors import DocTrPostProcess
+from .result import DocTrResult
+
+
+class WarpPredictor(BasicPredictor):
+    """WarpPredictor that inherits from BasicPredictor."""
+
+    entities = MODELS
+
+    def __init__(self, *args: List, **kwargs: Dict) -> None:
+        """Initializes WarpPredictor.
+
+        Args:
+            *args: Arbitrary positional arguments passed to the superclass.
+            **kwargs: Arbitrary keyword arguments passed to the superclass.
+        """
+        super().__init__(*args, **kwargs)
+        self.preprocessors, self.infer, self.postprocessors = self._build()
+
+    def _build_batch_sampler(self) -> ImageBatchSampler:
+        """Builds and returns an ImageBatchSampler instance.
+
+        Returns:
+            ImageBatchSampler: An instance of ImageBatchSampler.
+        """
+        return ImageBatchSampler()
+
+    def _get_result_class(self) -> type:
+        """Returns the warpping result, DocTrResult.
+
+        Returns:
+            type: The DocTrResult.
+        """
+        return DocTrResult
+
+    def _build(self) -> Tuple:
+        """Build the preprocessors, inference engine, and postprocessors based on the configuration.
+
+        Returns:
+            tuple: A tuple containing the preprocessors, inference engine, and postprocessors.
+        """
+        preprocessors = {"Read": ReadImage(format="BGR")}
+        preprocessors["Normalize"] = Normalize(mean=0.0, std=1.0, scale=1.0 / 255)
+        preprocessors["ToCHW"] = ToCHWImage()
+        preprocessors["ToBatch"] = ToBatch()
+
+        infer = StaticInfer(
+            model_dir=self.model_dir,
+            model_prefix=self.MODEL_FILE_PREFIX,
+            option=self.pp_option,
+        )
+
+        postprocessors = {"DocTrPostProcess": DocTrPostProcess()}
+        return preprocessors, infer, postprocessors
+
+    def process(self, batch_data: List[Union[str, np.ndarray]]) -> Dict[str, Any]:
+        """
+        Process a batch of data through the preprocessing, inference, and postprocessing.
+
+        Args:
+            batch_data (List[Union[str, np.ndarray], ...]): A batch of input data (e.g., image file paths).
+
+        Returns:
+            dict: A dictionary containing the input path, raw image, class IDs, scores, and label names for every instance of the batch. Keys include 'input_path', 'input_img', 'class_ids', 'scores', and 'label_names'.
+        """
+        batch_raw_imgs = self.preprocessors["Read"](imgs=batch_data.instances)
+        batch_imgs = self.preprocessors["Normalize"](imgs=batch_raw_imgs)
+        batch_imgs = self.preprocessors["ToCHW"](imgs=batch_imgs)
+        x = self.preprocessors["ToBatch"](imgs=batch_imgs)
+        batch_preds = self.infer(x=x)
+        batch_warp_preds = self.postprocessors["DocTrPostProcess"](batch_preds)
+
+        return {
+            "input_path": batch_data.input_paths,
+            "page_index": batch_data.page_indexes,
+            "input_img": batch_raw_imgs,
+            "doctr_img": batch_warp_preds,
+        }

paddlex/inference/models/image_unwarping/processors.py

@@ -0,0 +1,88 @@
+# copyright (c) 2024 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+from typing import List, Union, Tuple
+
+
+class DocTrPostProcess:
+    """
+    Post-processing class for cropping regions from images (though currently only performs scaling and color channel adjustments).
+
+    Attributes:
+        scale (np.float32): A scaling factor to be applied to the image pixel values.
+            Defaults to 255.0 if not provided.
+
+    Methods:
+        __call__(imgs: List[Union[np.ndarray, Tuple[np.ndarray, ...]]]) -> List[np.ndarray]:
+            Call method to process a list of images.
+        doctr(pred: Union[np.ndarray, Tuple[np.ndarray, ...]]) -> np.ndarray:
+            Method to process a single image or a tuple/list containing an image.
+    """
+
+    def __init__(self, scale: Union[str, float, None] = None):
+        """
+        Initializes the DocTrPostProcess class with a scaling factor.
+
+        Args:
+            scale (Union[str, float, None]): A scaling factor for the image pixel values.
+                If a string is provided, it will be converted to a float. Defaults to 255.0.
+        """
+        super().__init__()
+        self.scale = (
+            np.float32(scale) if isinstance(scale, (str, float)) else np.float32(255.0)
+        )
+
+    def __call__(
+        self, imgs: List[Union[np.ndarray, Tuple[np.ndarray, ...]]]
+    ) -> List[np.ndarray]:
+        """
+        Processes a list of images using the `doctr` method.
+
+        Args:
+            imgs (List[Union[np.ndarray, Tuple[np.ndarray, ...]]]): A list of images to process.
+                Each image can be a numpy array or a tuple containing a numpy array.
+
+        Returns:
+            List[np.ndarray]: A list of processed images.
+        """
+        return [self.doctr(img) for img in imgs]
+
+    def doctr(self, pred: Union[np.ndarray, Tuple[np.ndarray, ...]]) -> np.ndarray:
+        """
+        Processes a single image.
+
+        Args:
+            pred (Union[np.ndarray, Tuple[np.ndarray, ...]]): The image to process, which can be
+                a numpy array or a tuple containing a numpy array. Only the first element is used if it's a tuple.
+
+        Returns:
+            np.ndarray: The processed image.
+
+        Raises:
+            AssertionError: If the input is not a numpy array.
+        """
+        if isinstance(pred, tuple):
+            im = pred[0]
+        else:
+            im = pred
+        assert isinstance(
+            im, np.ndarray
+        ), "Invalid input 'im' in DocTrPostProcess. Expected a numpy array."
+        im = im.squeeze()
+        im = im.transpose(1, 2, 0)
+        im *= self.scale
+        im = im[:, :, ::-1]
+        im = im.astype("uint8", copy=False)
+        return im

paddlex/inference/models/image_unwarping/result.py

@@ -0,0 +1,45 @@
+# copyright (c) 2024 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import numpy as np
+from ...common.result import BaseCVResult, StrMixin, JsonMixin
+
+
+class DocTrResult(BaseCVResult):
+    """
+    Result class for DocTr, encapsulating the output of a document image processing task.
+
+    Attributes:
+        (inherited from BaseCVResult): Any attributes defined in the base class.
+
+    Methods:
+        _to_img(self) -> np.ndarray:
+            Converts the stored image result to a numpy array.
+    """
+
+    def _to_img(self) -> np.ndarray:
+        result = np.array(self["doctr_img"])
+        return {"res": result}
+
+    def _to_str(self, *args, **kwargs):
+        data = copy.deepcopy(self)
+        data.pop("input_img")
+        data["doctr_img"] = "..."
+        return JsonMixin._to_str(data, *args, **kwargs)
+
+    def _to_json(self, *args, **kwargs):
+        data = copy.deepcopy(self)
+        data.pop("input_img")
+        return JsonMixin._to_json(data, *args, **kwargs)

paddlex/inference/models/instance_segmentation/__init__.py

@@ -0,0 +1,15 @@
+# copyright (c) 2024 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .predictor import InstanceSegPredictor

paddlex/inference/models/instance_segmentation/predictor.py

@@ -0,0 +1,210 @@
+# copyright (c) 2024 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Union, Dict, List, Tuple, Sequence, Optional
+import numpy as np
+
+from ....modules.instance_segmentation.model_list import MODELS
+from ...common.batch_sampler import ImageBatchSampler
+from ..common import StaticInfer
+from ..object_detection.processors import (
+    ReadImage,
+    ToBatch,
+)
+from .processors import InstanceSegPostProcess
+from ..object_detection import DetPredictor
+from .result import InstanceSegResult
+from ....utils import logging
+
+
+class InstanceSegPredictor(DetPredictor):
+    """InstanceSegPredictor that inherits from DetPredictor."""
+
+    entities = MODELS
+
+    def __init__(self, *args, threshold: Optional[float] = None, **kwargs):
+        """Initializes InstanceSegPredictor.
+        Args:
+            *args: Arbitrary positional arguments passed to the superclass.
+            threshold (Optional[float], optional): The threshold for filtering out low-confidence predictions.
+                Defaults to None, in which case will use default from the config file.
+            **kwargs: Arbitrary keyword arguments passed to the superclass.
+        """
+        super().__init__(*args, **kwargs)
+
+        self.model_names_only_supports_batchsize_of_one = {
+            "SOLOv2",
+            "PP-YOLOE_seg-S",
+            "Cascade-MaskRCNN-ResNet50-vd-SSLDv2-FPN",
+            "Cascade-MaskRCNN-ResNet50-FPN",
+        }
+        if self.model_name in self.model_names_only_supports_batchsize_of_one:
+            logging.warning(
+                f"Instance Segmentation Models: \"{', '.join(list(self.model_names_only_supports_batchsize_of_one))}\" only supports prediction with a batch_size of one, "
+                "if you set the predictor with a batch_size larger than one, no error will occur, however, it will actually inference with a batch_size of one, "
+                f"which will lead to a slower inference speed. You are now using {self.config['Global']['model_name']}."
+            )
+
+        self.threshold = threshold
+
+    def _get_result_class(self) -> type:
+        """Returns the result class, InstanceSegResult.
+
+        Returns:
+            type: The InstanceSegResult class.
+        """
+        return InstanceSegResult
+
+    def _build(self) -> Tuple:
+        """Build the preprocessors, inference engine, and postprocessors based on the configuration.
+
+        Returns:
+            tuple: A tuple containing the preprocessors, inference engine, and postprocessors.
+        """
+        # build preprocess ops
+        pre_ops = [ReadImage(format="RGB")]
+        for cfg in self.config["Preprocess"]:
+            tf_key = cfg["type"]
+            func = self._FUNC_MAP[tf_key]
+            cfg.pop("type")
+            args = cfg
+            op = func(self, **args) if args else func(self)
+            if op:
+                pre_ops.append(op)
+        pre_ops.append(self.build_to_batch())
+
+        # build infer
+        infer = StaticInfer(
+            model_dir=self.model_dir,
+            model_prefix=self.MODEL_FILE_PREFIX,
+            option=self.pp_option,
+        )
+
+        # build postprocess op
+        post_op = self.build_postprocess()
+
+        return pre_ops, infer, post_op
+
+    def build_to_batch(self):
+
+        ordered_required_keys = (
+            "img_size",
+            "img",
+            "scale_factors",
+        )
+
+        return ToBatch(ordered_required_keys=ordered_required_keys)
+
+    def process(self, batch_data: List[Any], threshold: Optional[float] = None):
+        """
+        Process a batch of data through the preprocessing, inference, and postprocessing.
+
+        Args:
+            batch_data (List[Union[str, np.ndarray], ...]): A batch of input data (e.g., image file paths).
+
+        Returns:
+            dict: A dictionary containing the input path, raw image, box and mask
+                for every instance of the batch. Keys include 'input_path', 'input_img', 'boxes' and 'masks'.
+        """
+        datas = batch_data.instances
+        # preprocess
+        for pre_op in self.pre_ops[:-1]:
+            datas = pre_op(datas)
+
+        # use `ToBatch` format batch inputs
+        batch_inputs = self.pre_ops[-1](datas)
+
+        # do infer
+        if self.model_name in self.model_names_only_supports_batchsize_of_one:
+            batch_preds = []
+            for i in range(batch_inputs[0].shape[0]):
+                batch_inputs_ = [
+                    batch_input_[i][None, ...] for batch_input_ in batch_inputs
+                ]
+                batch_pred_ = self.infer(batch_inputs_)
+                batch_preds.append(batch_pred_)
+        else:
+            batch_preds = self.infer(batch_inputs)
+
+        # process a batch of predictions into a list of single image result
+        preds_list = self._format_output(batch_preds)
+
+        # postprocess
+        boxes_masks = self.post_op(
+            preds_list, datas, threshold if threshold is not None else self.threshold
+        )
+
+        return {
+            "input_path": batch_data.input_paths,
+            "page_index": batch_data.page_indexes,
+            "input_img": [data["ori_img"] for data in datas],
+            "boxes": [result["boxes"] for result in boxes_masks],
+            "masks": [result["masks"] for result in boxes_masks],
+        }
+
+    def _format_output(self, pred: Sequence[Any]) -> List[dict]:
+        """
+        Transform batch outputs into a list of single image output.
+
+        Args:
+            pred (Sequence[Any]): The input predictions, which can be either a list of 3 or 4 elements.
+                - When len(pred) == 4, it is expected to be in the format [boxes, class_ids, scores, masks],
+                  compatible with SOLOv2 output.
+                - When len(pred) == 3, it is expected to be in the format [boxes, box_nums, masks],
+                  compatible with Instance Segmentation output.
+
+        Returns:
+            List[dict]: A list of dictionaries, each containing either 'class_id' and 'masks' (for SOLOv2),
+                or 'boxes' and 'masks' (for Instance Segmentation), or just 'boxes' if no masks are provided.
+        """
+        box_idx_start = 0
+        pred_box = []
+
+        if isinstance(pred[0], list) and len(pred[0]) == 4:
+            # Adapt to SOLOv2, which only support prediction with a batch_size of 1.
+            pred_class_id = [[pred_[1], pred_[2]] for pred_ in pred]
+            pred_mask = [pred_[3] for pred_ in pred]
+            return [
+                {
+                    "class_id": np.array(pred_class_id[i]),
+                    "masks": np.array(pred_mask[i]),
+                }
+                for i in range(len(pred_class_id))
+            ]
+        if isinstance(pred[0], list) and len(pred[0]) == 3:
+            # Adapt to PP-YOLOE_seg-S, which only support prediction with a batch_size of 1.
+            return [
+                {"boxes": np.array(pred[i][0]), "masks": np.array(pred[i][2])}
+                for i in range(len(pred))
+            ]
+
+        pred_mask = []
+        for idx in range(len(pred[1])):
+            np_boxes_num = pred[1][idx]
+            box_idx_end = box_idx_start + np_boxes_num
+            np_boxes = pred[0][box_idx_start:box_idx_end]
+            pred_box.append(np_boxes)
+            np_masks = pred[2][box_idx_start:box_idx_end]
+            pred_mask.append(np_masks)
+            box_idx_start = box_idx_end
+
+        return [
+            {"boxes": np.array(pred_box[i]), "masks": np.array(pred_mask[i])}
+            for i in range(len(pred_box))
+        ]
+
+    def build_postprocess(self):
+        return InstanceSegPostProcess(
+            threshold=self.config["draw_threshold"], labels=self.config["label_list"]
+        )

paddlex/inference/models/instance_segmentation/processors.py

@@ -0,0 +1,105 @@
+# copyright (c) 2024 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from typing import List, Sequence, Tuple, Union, Optional
+
+import numpy as np
+from ....utils import logging
+from ..object_detection.processors import restructured_boxes
+
+import cv2
+
+
+def extract_masks_from_boxes(boxes, masks):
+    """
+    Extracts the portion of each mask that is within the corresponding box.
+    """
+    new_masks = []
+
+    for i, box in enumerate(boxes):
+        x_min, y_min, x_max, y_max = box["coordinate"]
+        x_min, y_min, x_max, y_max = map(
+            lambda x: int(round(x)), [x_min, y_min, x_max, y_max]
+        )
+
+        cropped_mask = masks[i][y_min:y_max, x_min:x_max]
+        new_masks.append(cropped_mask)
+
+    return new_masks
+
+
+class InstanceSegPostProcess(object):
+    """Save Result Transform"""
+
+    def __init__(self, threshold=0.5, labels=None):
+        super().__init__()
+        self.threshold = threshold
+        self.labels = labels
+
+    def apply(self, masks, img_size, boxes=None, class_id=None, threshold=None):
+        """apply"""
+        if boxes is not None:
+            expect_boxes = (boxes[:, 1] > threshold) & (boxes[:, 0] > -1)
+            boxes = boxes[expect_boxes, :]
+            boxes = restructured_boxes(boxes, self.labels, img_size)
+            masks = masks[expect_boxes, :, :]
+            masks = extract_masks_from_boxes(boxes, masks)
+            result = {"boxes": boxes, "masks": masks}
+        else:
+            mask_info = []
+            class_id = [list(item) for item in zip(class_id[0], class_id[1])]
+
+            selected_masks = []
+            for i, info in enumerate(class_id):
+                label_id = int(info[0])
+                if info[1] < threshold:
+                    continue
+                mask_info.append(
+                    {
+                        "label": self.labels[label_id],
+                        "score": info[1],
+                        "class_id": label_id,
+                    }
+                )
+                selected_masks.append(masks[i])
+            result = {"boxes": mask_info, "masks": selected_masks}
+
+        return result
+
+    def __call__(
+        self,
+        batch_outputs: List[dict],
+        datas: List[dict],
+        threshold: Optional[float] = None,
+    ):
+        """Apply the post-processing to a batch of outputs.
+
+        Args:
+            batch_outputs (List[dict]): The list of detection outputs.
+            datas (List[dict]): The list of input data.
+            threshold: Optional[float]: object score threshold for postprocess.
+
+        Returns:
+            List[Boxes]: The list of post-processed detection boxes.
+        """
+        outputs = []
+        for data, output in zip(datas, batch_outputs):
+            boxes_masks = self.apply(
+                img_size=data["ori_img_size"],
+                **output,
+                threshold=threshold if threshold is not None else self.threshold
+            )
+            outputs.append(boxes_masks)
+        return outputs