inference-models 0.18.3__py3-none-any.whl

inference_models/models/resnet/resnet_classification_torch.py

@@ -0,0 +1,305 @@
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import timm
+import torch
+from torch import nn
+
+from inference_models import (
+    ClassificationModel,
+    ClassificationPrediction,
+    MultiLabelClassificationModel,
+    MultiLabelClassificationPrediction,
+)
+from inference_models.configuration import DEFAULT_DEVICE
+from inference_models.entities import ColorFormat
+from inference_models.errors import CorruptedModelPackageError
+from inference_models.models.common.model_packages import get_model_package_contents
+from inference_models.models.common.roboflow.model_packages import (
+    InferenceConfig,
+    ResizeMode,
+    parse_class_names_file,
+    parse_inference_config,
+)
+from inference_models.models.common.roboflow.pre_processing import (
+    pre_process_network_input,
+)
+
+
+class ResNetClassifier(nn.Module):
+
+    def __init__(self, backbone: nn.Module, softmax_fused: bool):
+        super().__init__()
+        self._backbone = backbone
+        self._softmax_fused = softmax_fused
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        results = self._backbone(x)
+        if not self._softmax_fused:
+            results = torch.nn.functional.softmax(results, dim=-1)
+        return results
+
+
+class ResNetForClassificationTorch(ClassificationModel[torch.Tensor, torch.Tensor]):
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_name_or_path: str,
+        device: torch.device = DEFAULT_DEVICE,
+        **kwargs,
+    ) -> "ResNetForClassificationTorch":
+        model_package_content = get_model_package_contents(
+            model_package_dir=model_name_or_path,
+            elements=[
+                "class_names.txt",
+                "inference_config.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,
+            },
+        )
+        if inference_config.model_initialization is None:
+            raise CorruptedModelPackageError(
+                message="Expected model initialization parameters not provided in inference config.",
+                help_url="https://todo",
+            )
+        num_classes = inference_config.model_initialization.get("num_classes")
+        model_name = inference_config.model_initialization.get("model_name")
+        if not isinstance(num_classes, int):
+            raise CorruptedModelPackageError(
+                message="Expected model initialization parameter `num_classes` not provided or in invalid format.",
+                help_url="https://todo",
+            )
+        if not isinstance(model_name, str):
+            raise CorruptedModelPackageError(
+                message="Expected model initialization parameter `model_name` not provided or in invalid format.",
+                help_url="https://todo",
+            )
+        if inference_config.post_processing.type != "softmax":
+            raise CorruptedModelPackageError(
+                message="Expected softmax to be the post-processing",
+                help_url="https://todo",
+            )
+        backbone = timm.create_model(
+            model_name,
+            pretrained=False,
+            num_classes=num_classes,
+        ).to(device)
+        state_dict = torch.load(
+            model_package_content["weights.pth"],
+            weights_only=True,
+            map_location=device,
+        )
+        backbone.load_state_dict(state_dict)
+        model = ResNetClassifier(
+            backbone=backbone,
+            softmax_fused=inference_config.post_processing.fused,
+        ).to(device)
+        return cls(
+            model=model.eval(),
+            inference_config=inference_config,
+            class_names=class_names,
+            device=device,
+        )
+
+    def __init__(
+        self,
+        model: ResNetClassifier,
+        inference_config: InferenceConfig,
+        class_names: List[str],
+        device: torch.device,
+    ):
+        self._model = model
+        self._inference_config = inference_config
+        self._class_names = class_names
+        self._device = device
+
+    @property
+    def class_names(self) -> List[str]:
+        return self._class_names
+
+    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,
+    ) -> torch.Tensor:
+        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,
+        )[0]
+
+    def forward(self, pre_processed_images: torch.Tensor, **kwargs) -> torch.Tensor:
+        with torch.inference_mode():
+            return self._model(pre_processed_images)
+
+    def post_process(
+        self,
+        model_results: torch.Tensor,
+        **kwargs,
+    ) -> ClassificationPrediction:
+        return ClassificationPrediction(
+            class_id=model_results.argmax(dim=-1),
+            confidence=model_results,
+        )
+
+
+class ResNetMultiLabelClassifier(nn.Module):
+
+    def __init__(self, backbone: nn.Module, sigmoid_fused: bool):
+        super().__init__()
+        self._backbone = backbone
+        self._sigmoid_fused = sigmoid_fused
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        results = self._backbone(x)
+        if not self._sigmoid_fused:
+            results = torch.nn.functional.sigmoid(results)
+        return results
+
+
+class ResNetForMultiLabelClassificationTorch(
+    MultiLabelClassificationModel[torch.Tensor, torch.Tensor]
+):
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_name_or_path: str,
+        device: torch.device = DEFAULT_DEVICE,
+        **kwargs,
+    ) -> "ResNetForMultiLabelClassificationTorch":
+        model_package_content = get_model_package_contents(
+            model_package_dir=model_name_or_path,
+            elements=[
+                "class_names.txt",
+                "inference_config.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,
+            },
+        )
+        if inference_config.model_initialization is None:
+            raise CorruptedModelPackageError(
+                message="Expected model initialization parameters not provided in inference config.",
+                help_url="https://todo",
+            )
+        num_classes = inference_config.model_initialization.get("num_classes")
+        model_name = inference_config.model_initialization.get("model_name")
+        if not isinstance(num_classes, int):
+            raise CorruptedModelPackageError(
+                message="Expected model initialization parameter `num_classes` not provided or in invalid format.",
+                help_url="https://todo",
+            )
+        if not isinstance(model_name, str):
+            raise CorruptedModelPackageError(
+                message="Expected model initialization parameter `model_name` not provided or in invalid format.",
+                help_url="https://todo",
+            )
+        if inference_config.post_processing.type != "sigmoid":
+            raise CorruptedModelPackageError(
+                message="Expected sigmoid to be the post-processing",
+                help_url="https://todo",
+            )
+        backbone = timm.create_model(
+            model_name,
+            pretrained=False,
+            num_classes=num_classes,
+        ).to(device)
+        state_dict = torch.load(
+            model_package_content["weights.pth"],
+            weights_only=True,
+            map_location=device,
+        )
+        backbone.load_state_dict(state_dict)
+        model = ResNetMultiLabelClassifier(
+            backbone=backbone,
+            sigmoid_fused=inference_config.post_processing.fused,
+        ).to(device)
+        return cls(
+            model=model.eval(),
+            inference_config=inference_config,
+            class_names=class_names,
+            device=device,
+        )
+
+    def __init__(
+        self,
+        model: ResNetMultiLabelClassifier,
+        inference_config: InferenceConfig,
+        class_names: List[str],
+        device: torch.device,
+    ):
+        self._model = model
+        self._inference_config = inference_config
+        self._class_names = class_names
+        self._device = device
+
+    @property
+    def class_names(self) -> List[str]:
+        return self._class_names
+
+    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,
+    ) -> torch.Tensor:
+        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,
+        )[0]
+
+    def forward(self, pre_processed_images: torch.Tensor, **kwargs) -> torch.Tensor:
+        with torch.inference_mode():
+            return self._model(pre_processed_images)
+
+    def post_process(
+        self,
+        model_results: torch.Tensor,
+        confidence: float = 0.5,
+        **kwargs,
+    ) -> List[MultiLabelClassificationPrediction]:
+        results = []
+        for batch_element_confidence in model_results:
+            predicted_classes = torch.argwhere(
+                batch_element_confidence >= confidence
+            ).squeeze(dim=-1)
+            results.append(
+                MultiLabelClassificationPrediction(
+                    class_ids=predicted_classes,
+                    confidence=batch_element_confidence,
+                )
+            )
+        return results

inference_models/models/resnet/resnet_classification_trt.py

@@ -0,0 +1,369 @@
+from threading import Lock
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from inference_models import (
+    ClassificationModel,
+    ClassificationPrediction,
+    MultiLabelClassificationModel,
+    MultiLabelClassificationPrediction,
+)
+from inference_models.configuration import DEFAULT_DEVICE
+from inference_models.entities import ColorFormat
+from inference_models.errors import (
+    CorruptedModelPackageError,
+    MissingDependencyError,
+    ModelRuntimeError,
+)
+from inference_models.models.base.types import PreprocessedInputs
+from inference_models.models.common.cuda import (
+    use_cuda_context,
+    use_primary_cuda_context,
+)
+from inference_models.models.common.model_packages import get_model_package_contents
+from inference_models.models.common.roboflow.model_packages import (
+    InferenceConfig,
+    ResizeMode,
+    TRTConfig,
+    parse_class_names_file,
+    parse_inference_config,
+    parse_trt_config,
+)
+from inference_models.models.common.roboflow.pre_processing import (
+    pre_process_network_input,
+)
+from inference_models.models.common.trt import (
+    get_engine_inputs_and_outputs,
+    infer_from_trt_engine,
+    load_model,
+)
+
+try:
+    import tensorrt as trt
+except ImportError as import_error:
+    raise MissingDependencyError(
+        message=f"Could not import YOLOv8 model with TRT backend - this error means that some additional dependencies "
+        f"are not installed in the environment. If you run the `inference-models` library directly in your Python "
+        f"program, make sure the following extras of the package are installed: `trt10` - installation can only "
+        f"succeed for Linux and Windows machines with Cuda 12 installed. Jetson devices, should have TRT 10.x "
+        f"installed for all builds with Jetpack 6. "
+        f"If you see this error using Roboflow infrastructure, make sure the service you use does support the model. "
+        f"You can also contact Roboflow to get support.",
+        help_url="https://todo",
+    ) from import_error
+
+try:
+    import pycuda.driver as cuda
+except ImportError as import_error:
+    raise MissingDependencyError(
+        message="TODO", help_url="https://todo"
+    ) from import_error
+
+
+class ResNetForClassificationTRT(ClassificationModel[torch.Tensor, torch.Tensor]):
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_name_or_path: str,
+        device: torch.device = DEFAULT_DEVICE,
+        engine_host_code_allowed: bool = False,
+        **kwargs,
+    ) -> "ResNetForClassificationTRT":
+        if device.type != "cuda":
+            raise ModelRuntimeError(
+                message=f"TRT engine only runs on CUDA device - {device} device detected.",
+                help_url="https://todo",
+            )
+        model_package_content = get_model_package_contents(
+            model_package_dir=model_name_or_path,
+            elements=[
+                "class_names.txt",
+                "inference_config.json",
+                "trt_config.json",
+                "engine.plan",
+            ],
+        )
+        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,
+            },
+        )
+        if inference_config.post_processing.type != "softmax":
+            raise CorruptedModelPackageError(
+                message="Expected Softmax to be the post-processing",
+                help_url="https://todo",
+            )
+        trt_config = parse_trt_config(
+            config_path=model_package_content["trt_config.json"]
+        )
+        cuda.init()
+        cuda_device = cuda.Device(device.index or 0)
+        with use_primary_cuda_context(cuda_device=cuda_device) as cuda_context:
+            engine = load_model(
+                model_path=model_package_content["engine.plan"],
+                engine_host_code_allowed=engine_host_code_allowed,
+            )
+            execution_context = engine.create_execution_context()
+        inputs, outputs = get_engine_inputs_and_outputs(engine=engine)
+        if len(inputs) != 1:
+            raise CorruptedModelPackageError(
+                message=f"Implementation assume single model input, found: {len(inputs)}.",
+                help_url="https://todo",
+            )
+        if len(outputs) != 1:
+            raise CorruptedModelPackageError(
+                message=f"Implementation assume single model output, found: {len(outputs)}.",
+                help_url="https://todo",
+            )
+        return cls(
+            engine=engine,
+            input_name=inputs[0],
+            output_name=outputs[0],
+            class_names=class_names,
+            inference_config=inference_config,
+            trt_config=trt_config,
+            device=device,
+            cuda_context=cuda_context,
+            execution_context=execution_context,
+        )
+
+    def __init__(
+        self,
+        engine: trt.ICudaEngine,
+        input_name: str,
+        output_name: str,
+        class_names: List[str],
+        inference_config: InferenceConfig,
+        trt_config: TRTConfig,
+        device: torch.device,
+        cuda_context: cuda.Context,
+        execution_context: trt.IExecutionContext,
+    ):
+        self._engine = engine
+        self._input_name = input_name
+        self._output_names = [output_name]
+        self._class_names = class_names
+        self._inference_config = inference_config
+        self._trt_config = trt_config
+        self._device = device
+        self._cuda_context = cuda_context
+        self._execution_context = execution_context
+        self._lock = Lock()
+
+    @property
+    def class_names(self) -> List[str]:
+        return self._class_names
+
+    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,
+    ) -> torch.Tensor:
+        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,
+        )[0]
+
+    def forward(
+        self, pre_processed_images: PreprocessedInputs, **kwargs
+    ) -> torch.Tensor:
+        with self._lock:
+            with use_cuda_context(context=self._cuda_context):
+                return infer_from_trt_engine(
+                    pre_processed_images=pre_processed_images,
+                    trt_config=self._trt_config,
+                    engine=self._engine,
+                    context=self._execution_context,
+                    device=self._device,
+                    input_name=self._input_name,
+                    outputs=self._output_names,
+                )[0]
+
+    def post_process(
+        self,
+        model_results: torch.Tensor,
+        **kwargs,
+    ) -> ClassificationPrediction:
+        if self._inference_config.post_processing.fused:
+            confidence = model_results
+        else:
+            confidence = torch.nn.functional.softmax(model_results, dim=-1)
+        return ClassificationPrediction(
+            class_id=confidence.argmax(dim=-1),
+            confidence=confidence,
+        )
+
+
+class ResNetForMultiLabelClassificationTRT(
+    MultiLabelClassificationModel[torch.Tensor, torch.Tensor]
+):
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_name_or_path: str,
+        device: torch.device = DEFAULT_DEVICE,
+        engine_host_code_allowed: bool = False,
+        **kwargs,
+    ) -> "ResNetForMultiLabelClassificationTRT":
+        if device.type != "cuda":
+            raise ModelRuntimeError(
+                message=f"TRT engine only runs on CUDA device - {device} device detected.",
+                help_url="https://todo",
+            )
+        model_package_content = get_model_package_contents(
+            model_package_dir=model_name_or_path,
+            elements=[
+                "class_names.txt",
+                "inference_config.json",
+                "trt_config.json",
+                "engine.plan",
+            ],
+        )
+        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,
+            },
+        )
+        if inference_config.post_processing.type != "sigmoid":
+            raise CorruptedModelPackageError(
+                message="Expected sigmoid to be the post-processing",
+                help_url="https://todo",
+            )
+        trt_config = parse_trt_config(
+            config_path=model_package_content["trt_config.json"]
+        )
+        cuda.init()
+        cuda_device = cuda.Device(device.index or 0)
+        with use_primary_cuda_context(cuda_device=cuda_device) as cuda_context:
+            engine = load_model(
+                model_path=model_package_content["engine.plan"],
+                engine_host_code_allowed=engine_host_code_allowed,
+            )
+            execution_context = engine.create_execution_context()
+        inputs, outputs = get_engine_inputs_and_outputs(engine=engine)
+        if len(inputs) != 1:
+            raise CorruptedModelPackageError(
+                message=f"Implementation assume single model input, found: {len(inputs)}.",
+                help_url="https://todo",
+            )
+        if len(outputs) != 1:
+            raise CorruptedModelPackageError(
+                message=f"Implementation assume single model output, found: {len(outputs)}.",
+                help_url="https://todo",
+            )
+        return cls(
+            engine=engine,
+            input_name=inputs[0],
+            output_name=outputs[0],
+            class_names=class_names,
+            inference_config=inference_config,
+            trt_config=trt_config,
+            device=device,
+            cuda_context=cuda_context,
+            execution_context=execution_context,
+        )
+
+    def __init__(
+        self,
+        engine: trt.ICudaEngine,
+        input_name: str,
+        output_name: str,
+        class_names: List[str],
+        inference_config: InferenceConfig,
+        trt_config: TRTConfig,
+        device: torch.device,
+        cuda_context: cuda.Context,
+        execution_context: trt.IExecutionContext,
+    ):
+        self._engine = engine
+        self._input_name = input_name
+        self._output_names = [output_name]
+        self._class_names = class_names
+        self._inference_config = inference_config
+        self._trt_config = trt_config
+        self._device = device
+        self._cuda_context = cuda_context
+        self._execution_context = execution_context
+        self._lock = Lock()
+
+    @property
+    def class_names(self) -> List[str]:
+        return self._class_names
+
+    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,
+    ) -> torch.Tensor:
+        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,
+        )[0]
+
+    def forward(
+        self, pre_processed_images: PreprocessedInputs, **kwargs
+    ) -> torch.Tensor:
+        with self._lock:
+            with use_cuda_context(context=self._cuda_context):
+                return infer_from_trt_engine(
+                    pre_processed_images=pre_processed_images,
+                    trt_config=self._trt_config,
+                    engine=self._engine,
+                    context=self._execution_context,
+                    device=self._device,
+                    input_name=self._input_name,
+                    outputs=self._output_names,
+                )[0]
+
+    def post_process(
+        self,
+        model_results: torch.Tensor,
+        confidence: float = 0.5,
+        **kwargs,
+    ) -> List[MultiLabelClassificationPrediction]:
+        if self._inference_config.post_processing.fused:
+            model_results = model_results
+        else:
+            model_results = torch.nn.functional.sigmoid(model_results)
+        results = []
+        for batch_element_confidence in model_results:
+            predicted_classes = torch.argwhere(
+                batch_element_confidence >= confidence
+            ).squeeze(dim=-1)
+            results.append(
+                MultiLabelClassificationPrediction(
+                    class_ids=predicted_classes,
+                    confidence=batch_element_confidence,
+                )
+            )
+        return results