feature/add torchserve detectron2

examples/object_detector/detectron2/README.md

@@ -0,0 +1,67 @@
+# Object Detection using TorchServe and Detectron2
+
+## Overview
+
+This folder leverages **TorchServe** to deploy a Detectron2-based object detection model using a custom handler. It provides scalable and efficient object detection capabilities with support for both CPU and GPU environments.
+
+---
+
+## Table of Contents
+
+1. [Pre-requirements](#pre-requirements)
+2. [Installation](#installation)
+3. [Usage](#usage)
+4. [Documentation](#documentation)
+5. [Contributors](#contributors)
+
+---
+
+## Pre-requirements
+
+- **Python 3.8 or higher** (tested on Python 3.10.15).
+
+---
+
+## Installation
+
+Follow these steps to set up the project:
+
+1. Clone the repository:
+
+   ```bash
+   git clone https://github.com/pytorch/serve.git
+   ```
+
+2. Make sure the terminal's current directory is set to the folder where this README file is located:
+
+   ```bash
+   cd serve/examples/object_detector/detectron2
+   ```
+
+3. Install dependencies:
+
+   ```bash
+   pip install -r requirements.txt
+   pip install git+https://github.com/facebookresearch/detectron2.git && pip install numpy==1.21.6
+   ```
+
+---
+
+## Usage
+
+Refer to the [Documentation](#documentation) for detailed usage instructions.
+
+---
+
+## Documentation
+
+For detailed information on using TorchServe and Detectron2 for object detection, refer to the documentation provided in the [Upstart Commerce Blog](https://upstartcommerce.com/optimizing-pytorch-model-serving-at-scale-with-torchserve/).
+
+---
+
+## Contributors
+
+- **[Muhammad Mudassar](https://github.com/Mudassar-MLE)**  
+  - [LinkedIn](https://www.linkedin.com/in/muhammad-mudassar-a65645192/)  
+  - [Email](mailto:[email protected])
+---

examples/object_detector/detectron2/detectron2-handler.py

@@ -0,0 +1,265 @@
+import io
+import json
+import time
+import torch
+import logging
+import numpy as np
+from os import path
+from detectron2.config import get_cfg
+from PIL import Image, UnidentifiedImageError
+from detectron2.engine import DefaultPredictor
+from detectron2.utils.logger import setup_logger
+try:
+    import pillow_heif
+    import pillow_avif
+    import pillow_jxl
+    # Register openers for extended formats
+    pillow_heif.register_heif_opener()
+    # For pillow_avif and pillow_jxl, openers are registered upon import
+except ImportError as e:
+    raise ImportError(
+        "Please install 'pillow-heif', 'pillow-avif', and 'pillow-jxl' to handle extended image formats. "
+        f"Missing package error: {e}"
+    )
+########################################################################################################################################
+setup_logger()
+logger = logging.getLogger(__name__)
+logging.basicConfig(level=logging.INFO)
+########################################################################################################################################
+class ModelHandler:
+    """
+    A base ModelHandler implementation for loading and running Detectron2 models with TorchServe.
+    Compatible with both CPU and GPU.
+    """
+    def __init__(self):
+        """
+        Initialize the ModelHandler instance.
+        """
+        self.error = None
+        self._context = None
+        self._batch_size = 0
+        self.initialized = False
+        self.predictor = None
+        self.model_file = "model.pth"
+        self.config_file = "config.yaml"
+        self.device = "cpu"
+        if torch.cuda.is_available():
+            self.device = "cuda"
+            logger.info("Using GPU for inference.")
+        else:
+            logger.info("Using CPU for inference.")
+
+    def initialize(self, context):
+        """
+        Load the model and initialize the predictor.
+        Args:
+            context (Context): Initial context contains model server system properties.
+        """
+        logger.info("Initializing model...")
+
+        self._context = context
+        self._batch_size = context.system_properties.get("batch_size", 1)
+        model_dir = context.system_properties.get("model_dir")
+        model_path = path.join(model_dir, self.model_file)
+        config_path = path.join(model_dir, self.config_file)
+        logger.debug(f"Checking model file: {model_path} exists: {path.exists(model_path)}")
+        logger.debug(f"Checking config file: {config_path} exists: {path.exists(config_path)}")
+        if not path.exists(model_path):
+            error_msg = f"Model file {model_path} does not exist."
+            logger.error(error_msg)
+            self.error = error_msg
+            self.initialized = False
+            return
+        if not path.exists(config_path):
+            error_msg = f"Config file {config_path} does not exist."
+            logger.error(error_msg)
+            self.error = error_msg
+            self.initialized = False
+            return
+        try:
+            cfg = get_cfg()
+            cfg.merge_from_file(config_path)
+            cfg.MODEL.WEIGHTS = model_path
+            cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
+            cfg.MODEL.DEVICE = self.device
+            self.predictor = DefaultPredictor(cfg)
+            logger.info("Predictor initialized successfully.")
+            if self.predictor is None:
+                raise RuntimeError("Predictor initialization failed, the predictor is None.")
+            self.initialized = True
+            logger.info("Model initialization complete.")
+        except Exception as e:
+            error_msg = "Error during model initialization"
+            logger.exception(error_msg)
+            self.error = str(e)
+            self.initialized = False
+
+    def preprocess(self, batch):
+        """
+        Transform raw input into model input data.
+
+        Args:
+            batch (List[Dict]): List of raw requests, should match batch size.
+
+        Returns:
+            List[np.ndarray]: List of preprocessed images.
+        """
+        logger.info(f"Pre-processing started for a batch of {len(batch)}.")
+
+        images = []
+        for idx, request in enumerate(batch):
+            request_body = request.get("body")
+            if request_body is None:
+                error_msg = f"Request {idx} does not contain 'body'."
+                logger.error(error_msg)
+                raise ValueError(error_msg)
+            try:
+                image_stream = io.BytesIO(request_body)
+                try:
+                    pil_image = Image.open(image_stream)
+                    pil_image = pil_image.convert("RGB")
+                    img = np.array(pil_image)
+                    img = img[:, :, ::-1]
+                except UnidentifiedImageError as e:
+                    error_msg = f"Failed to identify image for request {idx}. Error: {e}"
+                    logger.error(error_msg)
+                    raise ValueError(error_msg)
+                except Exception as e:
+                    error_msg = f"Failed to decode image for request {idx}. Error: {e}"
+                    logger.error(error_msg)
+                    raise ValueError(error_msg)
+                images.append(img)
+            except Exception as e:
+                logger.exception(f"Error preprocessing request {idx}")
+                raise e
+        logger.info(f"Pre-processing finished for a batch of {len(batch)}.")
+        return images
+
+    def inference(self, model_input):
+        """
+        Perform inference on the model input.
+
+        Args:
+            model_input (List[np.ndarray]): List of preprocessed images.
+
+        Returns:
+            List[Dict]: List of inference outputs.
+        """
+        logger.info(f"Inference started for a batch of {len(model_input)}.")
+
+        outputs = []
+        for idx, image in enumerate(model_input):
+            try:
+                logger.debug(f"Processing image {idx}: shape={image.shape}, dtype={image.dtype}")
+                output = self.predictor(image)
+                outputs.append(output)
+            except Exception as e:
+                logger.exception(f"Error during inference on image {idx}")
+                raise e
+        logger.info(f"Inference finished for a batch of {len(model_input)}.")
+        return outputs
+
+    def postprocess(self, inference_outputs):
+        """
+        Post-process the inference outputs to a serializable format.
+
+        Args:
+            inference_outputs (List[Dict]): List of inference outputs.
+
+        Returns:
+            List[str]: List of JSON strings containing predictions.
+        """
+        start_time = time.time()
+        logger.info(f"Post-processing started at {start_time} for a batch of {len(inference_outputs)}.")
+        responses = []
+        for idx, output in enumerate(inference_outputs):
+            try:
+                predictions = output["instances"].to("cpu")
+                logger.debug(f"Available prediction fields: {predictions.get_fields().keys()}")
+                response = {}
+                if predictions.has("pred_classes"):
+                    classes = predictions.pred_classes.numpy().tolist()
+                    response["classes"] = classes
+                if predictions.has("pred_boxes"):
+                    boxes = predictions.pred_boxes.tensor.numpy().tolist()
+                    response["boxes"] = boxes
+                if predictions.has("scores"):
+                    scores = predictions.scores.numpy().tolist()
+                    response["scores"] = scores
+                if predictions.has("pred_masks"):
+                    response["masks_present"] = True
+                responses.append(json.dumps(response))
+            except Exception as e:
+                logger.exception(f"Error during post-processing of output {idx}")
+                raise e
+        elapsed_time = time.time() - start_time
+        logger.info(f"Post-processing finished for a batch of {len(inference_outputs)} in {elapsed_time:.2f} seconds.")
+
+        return responses
+
+    def handle(self, data, context):
+        """
+        Entry point for TorchServe to interact with the ModelHandler.
+
+        Args:
+            data (List[Dict]): Input data.
+            context (Context): Model server context.
+
+        Returns:
+            List[str]: List of predictions.
+        """
+        logger.info("Handling request...")
+        start_time = time.time()
+        if not self.initialized:
+            self.initialize(context)
+            if not self.initialized:
+                error_message = f"Model failed to initialize: {self.error}"
+                logger.error(error_message)
+                return [error_message]
+
+        if data is None:
+            error_message = "No data received for inference."
+            logger.error(error_message)
+            return [error_message]
+
+        try:
+            preprocess_start = time.time()
+            model_input = self.preprocess(data)
+            preprocess_time = time.time() - preprocess_start
+
+            inference_start = time.time()
+            model_output = self.inference(model_input)
+            inference_time = time.time() - inference_start
+
+            postprocess_start = time.time()
+            output = self.postprocess(model_output)
+            postprocess_time = time.time() - postprocess_start
+
+            total_time = time.time() - start_time
+            logger.info(
+                f"Handling request finished in {total_time:.2f} seconds. "
+                f"(Preprocess: {preprocess_time:.2f}s, "
+                f"Inference: {inference_time:.2f}s, "
+                f"Postprocess: {postprocess_time:.2f}s)"
+            )
+            return output
+        except Exception as e:
+            error_message = f"Error in handling request: {str(e)}"
+            logger.exception(error_message)
+            return [error_message]
+########################################################################################################################################
+_service = ModelHandler()
+
+def handle(data, context):
+    """
+    Entry point for TorchServe to interact with the ModelHandler.
+
+    Args:
+        data (List[Dict]): Input data.
+        context (Context): Model server context.
+
+    Returns:
+        List[str]: List of predictions.
+    """
+    return _service.handle(data, context)
+########################################################################################################################################

examples/object_detector/detectron2/requirements.txt

@@ -0,0 +1,14 @@
+opencv-python==4.10.0.84
+python-multipart==0.0.9
+torch==2.2.0
+torchvision==0.17.0
+transformers==4.44.2
+torchvision==0.17.0
+numpy==1.24.4
+torchserve==0.12.0
+torch-model-archiver==0.12.0
+torch-workflow-archiver==0.2.15
+pillow==11.0.0
+pillow-avif-plugin==1.4.6
+pillow-jxl-plugin==1.2.8
+pillow_heif==0.20.0