模型复现

GG-CNN

git地址：GitHub - dougsm/ggcnn: Generative Grasping CNN from "Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach" (RSS 2018)

问题1：libGL.so.1: cannot open shared object file: No such file or directory

   环境中缺少libGL.so.1文件，缺少GL库文件。

办法：更新apt-get：

apt-get update

   安装GL库：

apt install libgl1-mesa-glx

问题2:安装包

 pip install -r requirements.txt

问题3：引用数据指的是多个项目引用一份数据集

         可以使用绝对路径配置数据集的路径。

问题4：TiffPage 0: <COMPRESSION.LZW: 5> requires the 'imagecodecs' package

    读入.tif文件出错；

pip install imagecodecs-lite

问题5：ggcnn生成.tiff过程中无响应

 python -m utils.dataset_processing.generate_cornell_depth <Path To Dataset>

将“*”通配符去掉，原因是原文为了适配路径

pcds = glob.glob(os.path.join(args.path, 'pcd*[0-9].txt'))

 

 问题六配置训练的shell脚本

#!/bin/bash

 3 #使用的是GPU：0
 4 export CUDA_VISIBLE_DEVICES=0
 6 python3.6 -u train_ggcnn.py \
 7   --description training_example2 \
 8   #运行网络
 9   --network ggcnn2 \
10   #运行的數據集
11   --dataset jacquard \
12   #数据集
13   --dataset-path /opt/data/private/robot/jacquard
15 python3.6 -u train_ggcnn.py \
16   --description training_example \
17   #运行网络
18   --network ggcnn \
19   #运行的數據集
20   --dataset cornell \
21   #数据集
22   --dataset-path /opt/data/private/robot/cornell

结果显示：
----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1         [-1, 16, 300, 300]           1,952
              ReLU-2         [-1, 16, 300, 300]               0
            Conv2d-3         [-1, 16, 300, 300]           6,416
              ReLU-4         [-1, 16, 300, 300]               0
         MaxPool2d-5         [-1, 16, 150, 150]               0
            Conv2d-6         [-1, 16, 150, 150]           6,416
              ReLU-7         [-1, 16, 150, 150]               0
            Conv2d-8         [-1, 16, 150, 150]           6,416
              ReLU-9         [-1, 16, 150, 150]               0
        MaxPool2d-10           [-1, 16, 75, 75]               0
           Conv2d-11           [-1, 32, 75, 75]          12,832
             ReLU-12           [-1, 32, 75, 75]               0
           Conv2d-13           [-1, 32, 75, 75]          25,632
             ReLU-14           [-1, 32, 75, 75]               0
UpsamplingBilinear2d-15         [-1, 32, 150, 150]               0
           Conv2d-16         [-1, 16, 150, 150]           4,624
             ReLU-17         [-1, 16, 150, 150]               0
UpsamplingBilinear2d-18         [-1, 16, 300, 300]               0
           Conv2d-19         [-1, 16, 300, 300]           2,320
             ReLU-20         [-1, 16, 300, 300]               0
           Conv2d-21          [-1, 1, 300, 300]              17
           Conv2d-22          [-1, 1, 300, 300]              17
           Conv2d-23          [-1, 1, 300, 300]              17
           Conv2d-24          [-1, 1, 300, 300]              17
================================================================
Total params: 66,676
Trainable params: 66,676
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.34
Forward/backward pass size (MB): 110.55
Params size (MB): 0.25
Estimated Total Size (MB): 111.15
---------------------------------------------------------------

 训练集得到的结果：

225/249 = 0.903614

验证数据集：

IOU Results: 4570/5449 = 0.838686