GRCNN使用onnxruntime和tensorrt推理

下载GRCNN项目：https://github.com/skumra/robotic-grasping.git
导出onnx模型：

import torch

net = torch.load("trained-models/jacquard-rgbd-grconvnet3-drop0-ch32/epoch_42_iou_0.93")
x = torch.rand(1, 4, 300, 300).cuda()
torch.onnx.export(net, x,  "./grcnn.onnx", opset_version = 13)

onnx模型结构如下：

onnxruntime推理

import cv2
import onnxruntime
import numpy as np
from skimage.feature import peak_local_max


def process_data(rgb, depth, width, height, output_size):
    left = (width - output_size) // 2 
    top = (height - output_size) // 2
    right = (width + output_size) // 2 
    bottom = (height + output_size) // 2

    depth_img = depth[top:bottom, left:right]
    depth_img = np.clip((depth_img - depth_img.mean()), -1, 1)
    depth_img = depth_img.transpose(2, 0, 1)

    rgb_img = rgb[top:bottom, left:right]
    rgb_img = rgb_img.astype(np.float32) / 255.0
    rgb_img -= rgb_img.mean()
    rgb_img = rgb_img.transpose(2, 0, 1) 
    ret = np.concatenate((np.expand_dims(depth_img, 0), np.expand_dims(rgb_img, 0)), axis=1)
    return np.concatenate((np.expand_dims(depth_img, 0), np.expand_dims(rgb_img, 0)), axis=1)


if __name__ == '__main__':
    rgb = cv2.imread(