部署基于 TensorRT 量化的 YOLOV5s 4.0 模型

//量化预处理与训练保持一致，数据对齐
def preprocess_v1(image_raw):
    h, w, c = image_raw.shape
    image = cv2.cvtColor(image_raw, cv2.COLOR_BGR2RGB)
    # Calculate widht and height and paddings
    r_w = width / w
    r_h = height / h
    if r_h > r_w:
        tw = width
        th = int(r_w * h)
        tx1 = tx2 = 0
        ty1 = int((height - th) / 2)
        ty2 = height - th - ty1
    else:
        tw = int(r_h * w)
        th = height
        tx1 = int((width - tw) / 2)
        tx2 = width - tw - tx1
        ty1 = ty2 = 0
    # Resize the image with long side while maintaining ratio
    image = cv2.resize(image, (tw, th))
    # Pad the short side with (128,128,128)
    image = cv2.copyMakeBorder(
        image, ty1, ty2, tx1, tx2, cv2.BORDER_CONSTANT, (128, 128, 128)
    )
    image = image.astype(np.float32)
    # Normalize to [0,1]
    image /= 255.0
    # HWC to CHW format:
    image = np.transpose(image, [2, 0, 1])
    # CHW to NCHW format
    #image = np.expand_dims(image, axis=0)
    # Convert the image to row-major order, also known as "C order":
    #image = np.ascontiguousarray(image)
    return image

//构建IInt8EntropyCalibrator量化器
class Calibrator(trt.IInt8EntropyCalibrator):
    def __init__(self, stream, cache_file=""):
        trt.IInt8EntropyCalibrator.__init__(self)       
        self.stream = stream
        self.d_input = cuda.mem_alloc(self.stream.calibration_data.nbytes)
        self.cache_file = cache_file
        stream.reset()

    def get_batch_size(self):
        return self.stream.batch_size

    def get_batch(self, names):
        batch = self.stream.next_batch()
        if not batch.size:   
            return None

        cuda.memcpy_htod(self.d_input, batch)

        return [int(self.d_input)]

    def read_calibration_cache(self):
        # If there is a cache, use it instead of calibrating again. Otherwise, implicitly return None.
        if os.path.exists(self.cache_file):
            with open(self.cache_file, "rb") as f:
                logger.info("Using calibration cache to save time: {:}".format(self.cache_file))
                return f.read()

    def write_calibration_cache(self, cache):
        with open(self.cache_file, "wb") as f:
            logger.info("Caching calibration data for future use: {:}".format(self.cache_file))
            f.write(cache)

//加载onnx模型，构建tensorrt engine
def get_engine(max_batch_size=1, onnx_file_path="", engine_file_path="",\
               fp16_mode=False, int8_mode=False, calibration_stream=None, calibration_table_path="", save_engine=False):
    """Attempts to load a serialized engine if available, otherwise builds a new TensorRT engine and saves it."""
    def build_engine(max_batch_size, save_engine):
        """Takes an ONNX file and creates a TensorRT engine to run inference with"""
        with trt.Builder(TRT_LOGGER) as builder, \
                builder.create_network(1) as network,\
                trt.OnnxParser(network, TRT_LOGGER) as parser:
            
            # parse onnx model file
            if not os.path.exists(onnx_file_path):
                quit('ONNX file {} not found'.format(onnx_file_path))
            print('Loading ONNX file from path {}...'.format(onnx_file_path))
            with open(onnx_file_path, 'rb') as model:
                print('Beginning ONNX file parsing')
                parser.parse(model.read())
                assert network.num_layers > 0, 'Failed to parse ONNX model. \
                            Please check if the ONNX model is compatible '
            print('Completed parsing of ONNX file')
            print('Building an engine from file {}; this may take a while...'.format(onnx_file_path))        
            
            # build trt engine
            builder.max_batch_size = max_batch_size
            builder.max_workspace_size = 1 << 30 # 1GB
            builder.fp16_mode = fp16_mode
            if int8_mode:
                builder.int8_mode = int8_mode
                assert calibration_stream, 'Error: a calibration_stream should be provided for int8 mode'
                builder.int8_calibrator  = Calibrator(calibration_stream, calibration_table_path)
                print('Int8 mode enabled')
            engine = builder.build_cuda_engine(network) 
            if engine is None:
                print('Failed to create the engine')
                return None   
            print("Completed creating the engine")
            if save_engine:
                with open(engine_file_path, "wb") as f:
                    f.write(engine.serialize())
            return engine
        
    if os.path.exists(engine_file_path):
        # If a serialized engine exists, load it instead of building a new one.
        print("Reading engine from file {}".format(engine_file_path))
        with open(engine_file_path, "rb") as f, trt.Runtime(TRT_LOGGER) as runtime:
            return runtime.deserialize_cuda_engine(f.read())
    else:
        return build_engine(max_batch_size, save_engine)