博客
关于我
[TensorRT] ERROR: Network must have at least one output
阅读量:155 次
发布时间:2019-02-26

本文共 3909 字,大约阅读时间需要 13 分钟。

 

这个可以,rt6

生成rt,貌似可以

预测:

这个没试:

 

No module named 'pycuda'

 

 

[TensorRT] ERROR: Network must have at least one output

 

yolov3 转 tensorrt,运行onnx转tensorrt 有时会遇到上述错误。

onnx转tensorrt.py文件中,在parser.parse下with循环外添加两行代码

last_layer = network.get_layer(network.num_layers - 1)

network.mark_output(last_layer.get_output(0))
 

 

import loggingimport mathimport osimport pickleimport timeimport cv2import numpy as npimport tensorrt as trtimport torchfrom torchvision import transformslogger = logging.getLogger(__name__)TRT_LOGGER = trt.Logger()  # required by TensorRT def build_engine(onnx_file_path, engine_file_path, precision, max_batch_size, cache_file=None):    """Builds a new TensorRT engine and saves it, if no engine presents"""    if os.path.exists(engine_file_path):        logger.info('{} TensorRT engine already exists. Skip building engine...'.format(precision))        return    logger.info('Building {} TensorRT engine from onnx file...'.format(precision))    with trt.Builder(TRT_LOGGER) as b, b.create_network() as n, trt.OnnxParser(n, TRT_LOGGER) as p:        b.max_workspace_size = 1 << 30  # 1GB        b.max_batch_size = max_batch_size        if precision == 'fp16':            b.fp16_mode = True        elif precision == 'int8':            from ..calibrator import Calibrator            b.int8_mode = True            b.int8_calibrator = Calibrator(cache_file=cache_file)        elif precision == 'fp32':            pass        else:            logger.error('Engine precision not supported: {}'.format(precision))            raise NotImplementedError        # Parse model file        with open(onnx_file_path, 'rb') as model:            p.parse(model.read())        if p.num_errors:            logger.error('Parsing onnx file found {} errors.'.format(p.num_errors))        last_layer = network.get_layer(network.num_layers - 1)        network.mark_output(last_layer.get_output(0))        engine = b.build_cuda_engine(n)        print(engine_file_path)        with open(engine_file_path, "wb") as f:            f.write(engine.serialize())def add_coco_bbox(image, bbox, conf=1):    txt = '{}{:.1f}'.format('person', conf)    font = cv2.FONT_HERSHEY_SIMPLEX    cv2.rectangle(image, (bbox[0], bbox[1]), (bbox[2], bbox[3]), (0, 255, 255), 2)    cv2.putText(image, txt, (bbox[0], bbox[1] - 2),                font, 0.5, (0, 255, 0), thickness=1, lineType=cv2.LINE_AA)def add_coco_hp(image, points, keypoints_prob):    for j in range(5):        if keypoints_prob[j] > 0.5:            cv2.circle(image, (points[j, 0], points[j, 1]), 2, (255, 255, 0), -1)    return imageif __name__ == '__main__':    # 1. build trnsorrt engine    build_engine('mnist_cnn_3.onnx', 'mobile.trt', 'fp32', 1)    print('build trnsorrt engine done')    # 2. load trnsorrt engine    config = '../experiments/mobilenetv2_512x512.yaml'    body_engine = CenterNetTensorRTEngine(weight_file='../models/centerface/mobilenetv2-large/mobile.trt', config_file=config)    print('load trnsorrt engine done')    # 3. video for the tracking    cap = cv2.VideoCapture('/home/sai/YANG/image/video/zhaohang/01/01.mp4')    while (True):        # Capture frame-by-frame        ret, image = cap.read()        rgb_img = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)        start = time.clock()        detections = body_engine.run(image)[1]        print('time is:', time.clock() - start)        for i, bbox in enumerate(detections):            if bbox[4] > 0.7:                body_bbox = np.array(bbox[:4], dtype=np.int32)                body_prob = bbox[4]                add_coco_bbox(image, body_bbox, body_prob)                body_pose = np.array(bbox[5:15], dtype=np.int32)                keypoints = np.array(body_pose, dtype=np.int32).reshape(5, 2)                keypoints_prob = bbox[15:]                image = add_coco_hp(image, keypoints, keypoints_prob)        cv2.imshow('image result', image)        if cv2.waitKey(1) & 0xFF == ord('q'):            break

 

转载地址:http://rlzk.baihongyu.com/

你可能感兴趣的文章
MySQL 中文问题
查看>>
MySQL 中日志的面试题总结
查看>>
mysql 中的all,5分钟了解MySQL5.7中union all用法的黑科技
查看>>
MySQL 中的外键检查设置:SET FOREIGN_KEY_CHECKS = 1
查看>>
Mysql 中的日期时间字符串查询
查看>>
mysql 中索引的问题
查看>>
MySQL 中锁的面试题总结
查看>>
MySQL 中随机抽样:order by rand limit 的替代方案
查看>>
MySQL 为什么需要两阶段提交?
查看>>
mysql 为某个字段的值加前缀、去掉前缀
查看>>
mysql 主从
查看>>
mysql 主从 lock_mysql 主从同步权限mysql 行锁的实现
查看>>
mysql 主从互备份_mysql互为主从实战设置详解及自动化备份(Centos7.2)
查看>>
mysql 主从关系切换
查看>>
MYSQL 主从同步文档的大坑
查看>>
mysql 主键重复则覆盖_数据库主键不能重复
查看>>
Mysql 事务知识点与优化建议
查看>>
Mysql 优化 or
查看>>
mysql 优化器 key_mysql – 选择*和查询优化器
查看>>
MySQL 优化:Explain 执行计划详解
查看>>