yolo-world开放词汇检测onnxruntime和tensorrt推理

onnxruntime推理

导出onnx模型：

from ultralytics import YOLOWorld,YOLO
model = YOLOWorld('yolov8s-worldv2.pt')
model.set_classes(["glasses","shoes"])
model.save("yolo_world.pt")
model = YOLO('yolo_world.pt')
model.export(format="onnx", opset=12)

onnx模型结构如下图所示：

推理代码：

import cv2
import math
import numpy as np
import onnxruntime


class_names = ["glasses","shoes"]  
input_shape = (640, 640) 
score_threshold = 0.1 
nms_threshold = 0.5


def nms(boxes, scores, score_threshold, nms_threshold):
    x1 = boxes[:, 0]
    y1 = boxes[:, 1]
    x2 = boxes[:, 2]
    y2 = boxes[:, 3]
    areas = (y2 - y1 + 1) * (x2 - x1 + 1)
    keep = []
    index = scores.argsort()[::-1] 

    while index.size > 0:
        i = index[0]
        keep.append(i)
        x11 = np.maximum(x1[i], x1[index[1:]]) 
        y11 = np.maximum(y1[i], y1[index[1:]])
        x22 = np.minimum(x2[i], x2[index[1:]])
        y22 = np.minimum(y2[i], y2[index[1:]])
        w = np.maximum(0, x22 - x11 + 1)                              
        h = np.maximum(0, y22 - y11 + 1) 
        overlaps = w * h
        ious = overlaps / (areas[i] + areas[index[1:]] - overlaps)
        idx = np.where(ious <= nms_threshold)[0]
        index = index[idx + 1]
    return keep


def xywh2xyxy(x):
    y = np.copy(x)
    y[:, 0] = x[:, 0] - x[:, 2] / 2
    y[:, 1] = x[:, 1] - x[:, 3] / 2
    y[:, 2] = x[:, 0] + x[:, 2] / 2
    y[:, 3] = x[:, 1] + x[:, 3] / 2
    return y


def filter_box(outputs): #过滤掉无用的框    
    outputs = np.squeeze(outputs)
     
    boxes = []
    scores = []
    class_ids = []
    classes_scores = outputs[4:(4+len(class_names)), ...]  
       
    for i in range(outputs.shape[1]):              
        class_id = np.argmax(classes_scores[...,i])
        score = classes_scores[class_id][i]
        if score > score_threshold:
            boxes.append(np.concatenate([outputs[:4, i], np.array([score, class_id])]))
            scores.append(score)
            class_ids.append(class_id) 
            
    boxes = np.array(boxes)
    boxes = xywh2xyxy(boxes)
    scores = np.array(scores)
    indices = nms(boxes, scores, score_threshold, nms_threshold) 
    boxes = boxes[indices]
    return boxes


def letterbox(im, new_shape=(416, 416), color=(114, 114, 114)):
    # Resize and pad image while meeting stride-multiple constraints
    shape = im.shape[:2]  # current shape [height, width]

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    
    # Compute padding
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))    
    dw, dh = (new_shape[1] - new_unpad[0])/2, (new_shape[0] - new_unpad[1])/2  # wh padding 
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    
    if shape[::-1] != new_unpad:  # resize
        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return im


def scale_boxes(boxes, output_shape):
    # Rescale boxes (xyxy) from self.input_shape to shape
    gain = min(input_shape[0] / output_shape[0], input_shape[1] / output_shape[1])  # gain  = old / new
    pad = (input_shape[1] - output_shape[1] * gain) / 2, (input_shape[0] - output_shape[0] * gain) / 2  # wh padding
    boxes[..., [0, 2]] -= pad[0]  # x padding
    boxes[..., [1, 3]] -= pad[1]  # y padding
    boxes[..., :4] /= gain
    boxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, output_shape[1])  # x1, x2
    boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, output_shape[0])  # y1, y2
    return boxes


def draw(image, box_data):
    box_data = scale_boxes(box_data, image.shape)
    boxes = box_data[...,:4].astype(np.int32) 
    scores = box_data[...,4]
    classes = box_data[...,5].astype(np.int32)
   
    for box, score, cl in zip(boxes, scores, classes):
        top, left, right, bottom = box
        cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 1)
        cv2.putText(image, '{0} {1:.2f}'.format(class_names[cl], score), (top, left), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 1)


if __name__=="__main__":
    image = cv2.imread('bus.jpg', -1)
    input = letterbox(image, input_shape)
    input = input[:, :, ::-1].transpose(2, 0, 1).astype(dtype=np.float32)  #BGR2RGB和HWC2CHW
    input = input / 255.0
    input_tensor = []
    input_tensor.append(input)
    
    onnx_session = onnxruntime.InferenceSession('yolo_world.onnx', providers=['CPUExecutionProvider', 'CUDAExecutionProvider'])
        
    input_name = []
    for node in onnx_session.get_inputs():
        input_name.append(node.name)

    output_name = []
    for node in onnx_session.get_outputs():
        output_name.append(node.name)

    inputs = {}
    for name in input_name:
        inputs[name] =  np.array(input_tensor)
  
    outputs = onnx_session.run(None, inputs)[0]
    
    boxes = filter_box(outputs)
    draw(image, boxes)
    cv2.imwrite('result.jpg', image)

tensorrt推理

通过trtexec转换onnx模型得到engine模型：

./trtexec.exe --onnx=yolo_world.onnx --saveEngine=yolo_world.engine 

推理代码：

import cv2
import math
import numpy as np
import tensorrt as trt
import pycuda.autoinit 
import pycuda.driver as cuda 


class_names = ["glasses","shoes"]  
input_shape = (640, 640) 
score_threshold = 0.1 
nms_threshold = 0.5


def nms(boxes, scores, score_threshold, nms_threshold):
    x1 = boxes[:, 0]
    y1 = boxes[:, 1]
    x2 = boxes[:, 2]
    y2 = boxes[:, 3]
    areas = (y2 - y1 + 1) * (x2 - x1 + 1)
    keep = []
    index = scores.argsort()[::-1] 

    while index.size > 0:
        i = index[0]
        keep.append(i)
        x11 = np.maximum(x1[i], x1[index[1:]]) 
        y11 = np.maximum(y1[i], y1[index[1:]])
        x22 = np.minimum(x2[i], x2[index[1:]])
        y22 = np.minimum(y2[i], y2[index[1:]])
        w = np.maximum(0, x22 - x11 + 1)                              
        h = np.maximum(0, y22 - y11 + 1) 
        overlaps = w * h
        ious = overlaps / (areas[i] + areas[index[1:]] - overlaps)
        idx = np.where(ious <= nms_threshold)[0]
        index = index[idx + 1]
    return keep


def xywh2xyxy(x):
    y = np.copy(x)
    y[:, 0] = x[:, 0] - x[:, 2] / 2
    y[:, 1] = x[:, 1] - x[:, 3] / 2
    y[:, 2] = x[:, 0] + x[:, 2] / 2
    y[:, 3] = x[:, 1] + x[:, 3] / 2
    return y


def filter_box(outputs): #过滤掉无用的框    
    outputs = np.squeeze(outputs)
     
    boxes = []
    scores = []
    class_ids = []
    classes_scores = outputs[4:(4+len(class_names)), ...]  
       
    for i in range(outputs.shape[1]):              
        class_id = np.argmax(classes_scores[...,i])
        score = classes_scores[class_id][i]
        if score > score_threshold:
            boxes.append(np.concatenate([outputs[:4, i], np.array([score, class_id])]))
            scores.append(score)
            class_ids.append(class_id) 
            
    boxes = np.array(boxes)
    boxes = xywh2xyxy(boxes)
    scores = np.array(scores)
    indices = nms(boxes, scores, score_threshold, nms_threshold) 
    boxes = boxes[indices]
    return boxes


def letterbox(im, new_shape=(416, 416), color=(114, 114, 114)):
    # Resize and pad image while meeting stride-multiple constraints
    shape = im.shape[:2]  # current shape [height, width]

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    
    # Compute padding
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))    
    dw, dh = (new_shape[1] - new_unpad[0])/2, (new_shape[0] - new_unpad[1])/2  # wh padding 
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    
    if shape[::-1] != new_unpad:  # resize
        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return im


def scale_boxes(boxes, output_shape):
    # Rescale boxes (xyxy) from self.input_shape to shape
    gain = min(input_shape[0] / output_shape[0], input_shape[1] / output_shape[1])  # gain  = old / new
    pad = (input_shape[1] - output_shape[1] * gain) / 2, (input_shape[0] - output_shape[0] * gain) / 2  # wh padding
    boxes[..., [0, 2]] -= pad[0]  # x padding
    boxes[..., [1, 3]] -= pad[1]  # y padding
    boxes[..., :4] /= gain
    boxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, output_shape[1])  # x1, x2
    boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, output_shape[0])  # y1, y2
    return boxes


def draw(image, box_data):
    box_data = scale_boxes(box_data, image.shape)
    boxes = box_data[...,:4].astype(np.int32) 
    scores = box_data[...,4]
    classes = box_data[...,5].astype(np.int32)
   
    for box, score, cl in zip(boxes, scores, classes):
        top, left, right, bottom = box
        cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 1)
        cv2.putText(image, '{0} {1:.2f}'.format(class_names[cl], score), (top, left), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 1)


if __name__=="__main__":
    logger = trt.Logger(trt.Logger.WARNING)
    with open("yolo_world.engine", "rb") as f, trt.Runtime(logger) as runtime:
        engine = runtime.deserialize_cuda_engine(f.read())
    context = engine.create_execution_context()
    inputs_host = cuda.pagelocked_empty(trt.volume(context.get_binding_shape(0)), dtype=np.float32)
    outputs_host = cuda.pagelocked_empty(trt.volume(context.get_binding_shape(1)), dtype=np.float32)
    inputs_device = cuda.mem_alloc(inputs_host.nbytes)
    outputs_device = cuda.mem_alloc(outputs_host.nbytes)
    stream = cuda.Stream()
    
    image = cv2.imread('bus.jpg', -1)
    input = letterbox(image, input_shape)
    input = input[:, :, ::-1].transpose(2, 0, 1).astype(dtype=np.float32)  #BGR2RGB和HWC2CHW
    input = input / 255.0
    input = np.expand_dims(input, axis=0)     
    np.copyto(inputs_host, input.ravel())

    with engine.create_execution_context() as context:
        cuda.memcpy_htod_async(inputs_device, inputs_host, stream)
        context.execute_async_v2(bindings=[int(inputs_device), int(outputs_device)], stream_handle=stream.handle)
        cuda.memcpy_dtoh_async(outputs_host, outputs_device, stream)
        stream.synchronize()  
        boxes = filter_box(outputs_host.reshape(context.get_binding_shape(1)))
        draw(image, boxes)
        cv2.imwrite('result.jpg', image)

推理结果可视化：