inference-models 0.18.3__py3-none-any.whl

inference_models/models/rfdetr/rfdetr_object_detection_pytorch.py

@@ -0,0 +1,470 @@
+import os.path
+from copy import deepcopy
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from inference_models import Detections, ObjectDetectionModel
+from inference_models.configuration import DEFAULT_DEVICE
+from inference_models.entities import ColorFormat
+from inference_models.errors import (
+    CorruptedModelPackageError,
+    ModelLoadingError,
+    ModelRuntimeError,
+)
+from inference_models.logger import LOGGER
+from inference_models.models.common.model_packages import get_model_package_contents
+from inference_models.models.common.roboflow.model_packages import (
+    ColorMode,
+    DivisiblePadding,
+    InferenceConfig,
+    NetworkInputDefinition,
+    PreProcessingMetadata,
+    ResizeMode,
+    TrainingInputSize,
+    parse_class_names_file,
+    parse_inference_config,
+)
+from inference_models.models.common.roboflow.pre_processing import (
+    pre_process_network_input,
+)
+from inference_models.models.rfdetr.class_remapping import (
+    ClassesReMapping,
+    prepare_class_remapping,
+)
+from inference_models.models.rfdetr.common import parse_model_type
+from inference_models.models.rfdetr.default_labels import resolve_labels
+from inference_models.models.rfdetr.post_processor import PostProcess
+from inference_models.models.rfdetr.rfdetr_base_pytorch import (
+    LWDETR,
+    RFDETRBaseConfig,
+    RFDETRLargeConfig,
+    RFDETRMediumConfig,
+    RFDETRNanoConfig,
+    RFDETRSmallConfig,
+    build_model,
+)
+
+try:
+    torch.set_float32_matmul_precision("high")
+except:
+    pass
+
+CONFIG_FOR_MODEL_TYPE = {
+    "rfdetr-nano": RFDETRNanoConfig,
+    "rfdetr-small": RFDETRSmallConfig,
+    "rfdetr-medium": RFDETRMediumConfig,
+    "rfdetr-base": RFDETRBaseConfig,
+    "rfdetr-large": RFDETRLargeConfig,
+}
+
+RESIZE_MODES_TO_REVERT_PADDING = {
+    ResizeMode.LETTERBOX,
+    ResizeMode.LETTERBOX_REFLECT_EDGES,
+}
+
+
+class RFDetrForObjectDetectionTorch(
+    (ObjectDetectionModel[torch.Tensor, PreProcessingMetadata, dict])
+):
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_name_or_path: str,
+        device: torch.device = DEFAULT_DEVICE,
+        model_type: Optional[str] = None,
+        labels: Optional[Union[str, List[str]]] = None,
+        resolution: Optional[int] = None,
+        **kwargs,
+    ) -> "RFDetrForObjectDetectionTorch":
+        if os.path.isfile(model_name_or_path):
+            return cls.from_checkpoint_file(
+                checkpoint_path=model_name_or_path,
+                model_type=model_type,
+                labels=labels,
+                resolution=resolution,
+            )
+        model_package_content = get_model_package_contents(
+            model_package_dir=model_name_or_path,
+            elements=[
+                "class_names.txt",
+                "inference_config.json",
+                "model_type.json",
+                "weights.pth",
+            ],
+        )
+        class_names = parse_class_names_file(
+            class_names_path=model_package_content["class_names.txt"]
+        )
+        inference_config = parse_inference_config(
+            config_path=model_package_content["inference_config.json"],
+            allowed_resize_modes={
+                ResizeMode.STRETCH_TO,
+                ResizeMode.LETTERBOX,
+                ResizeMode.CENTER_CROP,
+                ResizeMode.LETTERBOX_REFLECT_EDGES,
+            },
+        )
+        classes_re_mapping = None
+        if inference_config.class_names_operations:
+            class_names, classes_re_mapping = prepare_class_remapping(
+                class_names=class_names,
+                class_names_operations=inference_config.class_names_operations,
+                device=device,
+            )
+        weights_dict = torch.load(
+            model_package_content["weights.pth"],
+            map_location=device,
+            weights_only=False,
+        )["model"]
+        model_type = parse_model_type(
+            config_path=model_package_content["model_type.json"]
+        )
+        if model_type not in CONFIG_FOR_MODEL_TYPE:
+            raise CorruptedModelPackageError(
+                message=f"Model package describes model_type as '{model_type}' which is not supported. "
+                f"Supported model types: {list(CONFIG_FOR_MODEL_TYPE.keys())}.",
+                help_url="https://todo",
+            )
+        model_config = CONFIG_FOR_MODEL_TYPE[model_type](device=device)
+        checkpoint_num_classes = weights_dict["class_embed.bias"].shape[0]
+        model_config.num_classes = checkpoint_num_classes - 1
+        model_config.resolution = (
+            inference_config.network_input.training_input_size.height
+        )
+        model = build_model(config=model_config)
+        model.load_state_dict(weights_dict)
+        model = model.eval().to(device)
+        post_processor = PostProcess()
+        return cls(
+            model=model,
+            class_names=class_names,
+            classes_re_mapping=classes_re_mapping,
+            device=device,
+            inference_config=inference_config,
+            post_processor=post_processor,
+            resolution=model_config.resolution,
+        )
+
+    @classmethod
+    def from_checkpoint_file(
+        cls,
+        checkpoint_path: str,
+        model_type: Optional[str] = None,
+        labels: Optional[Union[str, List[str]]] = None,
+        resolution: Optional[int] = None,
+        device: torch.device = DEFAULT_DEVICE,
+    ):
+        if model_type is None:
+            raise ModelLoadingError(
+                message="While loading RFDetr model (using torch backend) could not determine `model_type`. "
+                "If you used `RFDetrForObjectDetectionTorch` directly imported in your code, please pass "
+                f"one of the value: {CONFIG_FOR_MODEL_TYPE.keys()} as the parameter. If you see this "
+                f"error, while using `AutoModel.from_pretrained(...)` or thrown from managed Roboflow service, "
+                f"this is a bug - raise the issue: https://github.com/roboflow/inference/issue providing "
+                f"full context.",
+                help_url="https://todo",
+            )
+        weights_dict = torch.load(
+            checkpoint_path,
+            map_location=device,
+            weights_only=False,
+        )["model"]
+        if model_type not in CONFIG_FOR_MODEL_TYPE:
+            raise ModelLoadingError(
+                message=f"Model package describes model_type as '{model_type}' which is not supported. "
+                f"Supported model types: {list(CONFIG_FOR_MODEL_TYPE.keys())}.",
+                help_url="https://todo",
+            )
+        model_config = CONFIG_FOR_MODEL_TYPE[model_type](device=device)
+        divisibility = model_config.num_windows * model_config.patch_size
+        if resolution is not None:
+            if resolution < 0 or resolution % divisibility != 0:
+                raise ModelLoadingError(
+                    message=f"Attempted to load RFDetr model (using torch backend) with `resolution` parameter which "
+                    f"is invalid - the model required positive value divisible by 56. Make sure you used "
+                    f"proper value, corresponding to the one used to train the model.",
+                    help_url="https://todo",
+                )
+            model_config.resolution = resolution
+        inference_config = InferenceConfig(
+            network_input=NetworkInputDefinition(
+                training_input_size=TrainingInputSize(
+                    height=model_config.resolution,
+                    width=model_config.resolution,
+                ),
+                dynamic_spatial_size_supported=True,
+                dynamic_spatial_size_mode=DivisiblePadding(
+                    type="pad-to-be-divisible",
+                    value=divisibility,
+                ),
+                color_mode=ColorMode.BGR,
+                resize_mode=ResizeMode.STRETCH_TO,
+                input_channels=3,
+                scaling_factor=255,
+                normalization=([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+            )
+        )
+        checkpoint_num_classes = weights_dict["class_embed.bias"].shape[0]
+        model_config.num_classes = checkpoint_num_classes - 1
+        model = build_model(config=model_config)
+        if labels is None:
+            class_names = [f"class_{i}" for i in range(checkpoint_num_classes)]
+        elif isinstance(labels, str):
+            class_names = resolve_labels(labels=labels)
+        else:
+            class_names = labels
+        if checkpoint_num_classes != len(class_names):
+            raise ModelLoadingError(
+                message=f"Checkpoint pointed to load RFDetr defines {checkpoint_num_classes} output classes, but "
+                f"loaded labels define {len(class_names)} classes - fix the value of `labels` parameter.",
+                help_url="https://todo",
+            )
+        model.load_state_dict(weights_dict)
+        model = model.eval().to(device)
+        post_processor = PostProcess()
+        return cls(
+            model=model,
+            class_names=class_names,
+            classes_re_mapping=None,
+            device=device,
+            inference_config=inference_config,
+            post_processor=post_processor,
+            resolution=model_config.resolution,
+        )
+
+    def __init__(
+        self,
+        model: LWDETR,
+        inference_config: InferenceConfig,
+        class_names: List[str],
+        classes_re_mapping: Optional[ClassesReMapping],
+        device: torch.device,
+        post_processor: PostProcess,
+        resolution: int,
+    ):
+        self._model = model
+        self._inference_config = inference_config
+        self._class_names = class_names
+        self._classes_re_mapping = classes_re_mapping
+        self._post_processor = post_processor
+        self._device = device
+        self._resolution = resolution
+        self._has_warned_about_not_being_optimized_for_inference = False
+        self._inference_model: Optional[LWDETR] = None
+        self._optimized_has_been_compiled = False
+        self._optimized_batch_size = None
+        self._optimized_dtype = None
+
+    @property
+    def class_names(self) -> List[str]:
+        return self._class_names
+
+    def optimize_for_inference(
+        self,
+        compile: bool = True,
+        batch_size: int = 1,
+        dtype: torch.dtype = torch.float32,
+    ) -> None:
+        self.remove_optimized_model()
+        self._inference_model = deepcopy(self._model)
+        self._inference_model.eval()
+        self._inference_model.export()
+        self._inference_model = self._inference_model.to(dtype=dtype)
+        self._optimized_dtype = dtype
+        if compile:
+            self._inference_model = torch.jit.trace(
+                self._inference_model,
+                torch.randn(
+                    batch_size,
+                    3,
+                    self._resolution,
+                    self._resolution,
+                    device=self._device,
+                    dtype=dtype,
+                ),
+            )
+            self._optimized_has_been_compiled = True
+            self._optimized_batch_size = batch_size
+
+    def remove_optimized_model(self) -> None:
+        self._has_warned_about_not_being_optimized_for_inference = False
+        self._inference_model = None
+        self._optimized_has_been_compiled = False
+        self._optimized_batch_size = None
+
+    def pre_process(
+        self,
+        images: Union[torch.Tensor, List[torch.Tensor], np.ndarray, List[np.ndarray]],
+        input_color_format: Optional[ColorFormat] = None,
+        image_size: Optional[Tuple[int, int]] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, List[PreProcessingMetadata]]:
+        return pre_process_network_input(
+            images=images,
+            image_pre_processing=self._inference_config.image_pre_processing,
+            network_input=self._inference_config.network_input,
+            target_device=self._device,
+            input_color_format=input_color_format,
+            image_size_wh=image_size,
+        )
+
+    def forward(self, pre_processed_images: torch.Tensor, **kwargs) -> dict:
+        if (
+            self._inference_model is None
+            and not self._has_warned_about_not_being_optimized_for_inference
+        ):
+            LOGGER.warning(
+                "Model is not optimized for inference. "
+                "Latency may be higher than expected. "
+                "You can optimize the model for inference by calling model.optimize_for_inference()."
+            )
+            self._has_warned_about_not_being_optimized_for_inference = True
+        if self._inference_model is not None:
+            if (self._resolution, self._resolution) != tuple(
+                pre_processed_images.shape[2:]
+            ):
+                raise ModelRuntimeError(
+                    message=f"Resolution mismatch. Model was optimized for resolution {self._resolution}, "
+                    f"but got {tuple(pre_processed_images.shape[2:])}. "
+                    "You can explicitly remove the optimized model by calling model.remove_optimized_model().",
+                    help_url="https://todo",
+                )
+            if self._optimized_has_been_compiled:
+                if self._optimized_batch_size != pre_processed_images.shape[0]:
+                    raise ModelRuntimeError(
+                        message="Batch size mismatch. Optimized model was compiled for batch size "
+                        f"{self._optimized_batch_size}, but got {pre_processed_images.shape[0]}. "
+                        "You can explicitly remove the optimized model by calling model.remove_optimized_model(). "
+                        "Alternatively, you can recompile the optimized model for a different batch size "
+                        "by calling model.optimize_for_inference(batch_size=<new_batch_size>).",
+                        help_url="https://todo",
+                    )
+        with torch.inference_mode():
+            if self._inference_model:
+                predictions = self._inference_model(
+                    pre_processed_images.to(dtype=self._optimized_dtype)
+                )
+            else:
+                predictions = self._model(pre_processed_images)
+            if isinstance(predictions, tuple):
+                predictions = {
+                    "pred_logits": predictions[1],
+                    "pred_boxes": predictions[0],
+                }
+            return predictions
+
+    def post_process(
+        self,
+        model_results: dict,
+        pre_processing_meta: List[PreProcessingMetadata],
+        threshold: float = 0.5,
+        **kwargs,
+    ) -> List[Detections]:
+        if (
+            self._inference_config.network_input.resize_mode
+            in RESIZE_MODES_TO_REVERT_PADDING
+        ):
+            un_padding_results = []
+            for out_box_tensor, image_metadata in zip(
+                model_results["pred_boxes"], pre_processing_meta
+            ):
+                box_center_offsets = torch.as_tensor(  # bboxes in format cxcywh now, so only cx, cy to be pushed
+                    [
+                        image_metadata.pad_left / image_metadata.inference_size.width,
+                        image_metadata.pad_top / image_metadata.inference_size.height,
+                        0.0,
+                        0.0,
+                    ],
+                    dtype=out_box_tensor.dtype,
+                    device=out_box_tensor.device,
+                )
+                ox_padding = (
+                    image_metadata.pad_left + image_metadata.pad_right
+                ) / image_metadata.inference_size.width
+                oy_padding = (
+                    image_metadata.pad_top + image_metadata.pad_bottom
+                ) / image_metadata.inference_size.height
+                box_wh_offsets = torch.as_tensor(  # bboxes in format cxcywh now, so only cx, cy to be pushed
+                    [
+                        1.0 - ox_padding,
+                        1.0 - oy_padding,
+                        1.0 - ox_padding,
+                        1.0 - oy_padding,
+                    ],
+                    dtype=out_box_tensor.dtype,
+                    device=out_box_tensor.device,
+                )
+                out_box_tensor = (out_box_tensor - box_center_offsets) / box_wh_offsets
+                un_padding_results.append(out_box_tensor)
+            model_results["pred_boxes"] = torch.stack(un_padding_results, dim=0)
+        if self._inference_config.network_input.resize_mode is ResizeMode.CENTER_CROP:
+            orig_sizes = [
+                (
+                    round(e.inference_size.height / e.scale_height),
+                    round(e.inference_size.width / e.scale_width),
+                )
+                for e in pre_processing_meta
+            ]
+        else:
+            orig_sizes = [
+                (e.size_after_pre_processing.height, e.size_after_pre_processing.width)
+                for e in pre_processing_meta
+            ]
+        target_sizes = torch.tensor(orig_sizes, device=self._device)
+        results = self._post_processor(model_results, target_sizes=target_sizes)
+        detections_list = []
+        for image_result, image_metadata in zip(results, pre_processing_meta):
+            scores = image_result["scores"]
+            labels = image_result["labels"]
+            boxes = image_result["boxes"]
+            if self._classes_re_mapping is not None:
+                remapping_mask = torch.isin(
+                    labels, self._classes_re_mapping.remaining_class_ids
+                )
+                scores = scores[remapping_mask]
+                labels = self._classes_re_mapping.class_mapping[labels[remapping_mask]]
+                boxes = boxes[remapping_mask]
+            keep = scores > threshold
+            scores = scores[keep]
+            labels = labels[keep]
+            boxes = boxes[keep]
+            if (
+                self._inference_config.network_input.resize_mode
+                is ResizeMode.CENTER_CROP
+            ):
+                offsets = torch.as_tensor(
+                    [
+                        image_metadata.pad_left,
+                        image_metadata.pad_top,
+                        image_metadata.pad_left,
+                        image_metadata.pad_top,
+                    ],
+                    dtype=boxes.dtype,
+                    device=boxes.device,
+                )
+                boxes[:, :4].sub_(offsets)
+            if (
+                image_metadata.static_crop_offset.offset_x != 0
+                or image_metadata.static_crop_offset.offset_y != 0
+            ):
+                static_crop_offsets = torch.as_tensor(
+                    [
+                        image_metadata.static_crop_offset.offset_x,
+                        image_metadata.static_crop_offset.offset_y,
+                        image_metadata.static_crop_offset.offset_x,
+                        image_metadata.static_crop_offset.offset_y,
+                    ],
+                    dtype=boxes.dtype,
+                    device=boxes.device,
+                )
+                boxes[:, :4].add_(static_crop_offsets)
+            detections = Detections(
+                xyxy=boxes.round().int(),
+                confidence=scores,
+                class_id=labels.int(),
+            )
+            detections_list.append(detections)
+        return detections_list