grasp_multiObject_multiGrasp论文代码复现遇到的问题

源代码地址：grasp_multiObject_multiGrasp

原作者的开发环境是tensorflow-gpu+python2

而我的开发环境是tensorflow-cpu+python3.....还没有装matlab

首先修改一些配置文件使程序能够在CPU环境下运行，参考https://blog.csdn.net/m0_38024766/article/details/90712715

下载好Cornell Grasp dataset之后，首先要将数据集分类，用到的是PNG和POS.txt，因为系统里没装matlab,就自己写了一个python程序代替./data/scripts/dataPreprocessingTest_fasterrcnn_split.m

import random
import cv2
import skimage.io as io
import numpy as np
import os
import math
list_idx=list(range(100,950))+list(range(1000,1035))
random.shuffle(list_idx)
# print(len(list_idx))
train_list=list_idx[171:]
test_list=list_idx[:171]
cropSize=227
pi=3.14159
def dataPreprocessing_fasterrcnn(imagein,bbsIn_all,cropSize,translationShiftNumber,roatateAngleNumber):
    imgCrop=imagein[65:414,145:494]
    imgpadding=cv2.copyMakeBorder(imgCrop,75,75,75,75,cv2.BORDER_CONSTANT,value=[0,0,0])
    count=1
    rows, cols, channel = imgpadding.shape #绕图像的中心旋转 #参数：旋转中心 旋转度数 scale M = cv2.getRotationMatrix2D((cols/2, rows/2), 30, 1) #参数：原始图像 旋转参数 元素图像宽高 rotated = cv2.warpAffine(src, M, (cols, rows))
    m,n=np.shape(bbsIn_all)
    bbsnum=m//4

    
    count=0
    imgout={}
    bbsout={}
    dev=0
    for i_rotate in range(roatateAngleNumber*translationShiftNumber*translationShiftNumber):
        theta=random.randint(0,360)
        dx=random.randint(0,101)-51
        dy=random.randint(0,101)-51
        M = cv2.getRotationMatrix2D((cols/2, rows/2), theta, 1)
        imgrotate = cv2.warpAffine(imgpadding, M, (cols, rows))
        imgcroprotate=imgrotate[int(rows/2-160-dy):int(rows/2+160-dy),int(cols/2-160-dx):int(cols/2+160-dx)]
        imgCropRotateresize=cv2.resize(imgcroprotate,(cropSize,cropSize))
        bbs=np.zeros((m,n))
        lst=[]
        
        for idx in range(bbsnum):
            bbsin=bbsIn_all[idx*4:idx*4+4]
                # print(bbsin)

            # print(bbsin)
            if np.isnan(bbsin).any():
                dev=1
                lst.extend([i for i in range(idx*4,idx*4+4)])
                continue
            bbsinshift=bbsin-[320,240]
                # print(bbsinshift)
            R=[[math.cos(theta/180*pi),-math.sin(theta/180*pi)],[math.sin(theta/180*pi),math.cos(theta/180*pi)]]
            bbsRotated=np.matmul(bbsinshift,R)
            bbsInShiftBack=(bbsRotated+[160,160]+[dx,dy])*cropSize/320
            bbs[idx*4:idx*4+4]=bbsInShiftBack
        imgout[count]=imgCropRotateresize
        if dev:
            bbs=np.delete(bbs,lst,axis=0)
            dev=0
        # print(bbs)
        bbsout[count]=bbs
        count+=1
    return imgout,bbsout





imgDataDir = '/home/sjy/grasp_multiObject_multiGrasp/data/grasps/image/'
posDataDir='/home/sjy/grasp_multiObject_multiGrasp/data/grasps/pos_label/'
# negDataDir= '/home/sjy/grasp_multiObject_multiGrasp/data/grasps/neg_label/'
imgDataOutDir='/home/sjy/grasp_data/data/Images/'
annotationDataOutDir='/home/sjy/grasp_data/data/Annotations/'
imgSetTrain = '/home/sjy/grasp_data/data/ImageSets/train1.txt'
imgSetTrain1 = '/home/sjy/grasp_data/data/ImageSets/train.txt'
imgSetTest = '/home/sjy/grasp_data/data/ImageSets/test.txt'
imgFiles=[x.name for x in os.scandir(imgDataDir)]
txtFiles=[x.name for x in os.scandir(posDataDir)]

# f=open(imgSetTest,"a")
for x in imgFiles:
    print(x)
    if int(x[3:7]) in test_list:
            f=open(imgSetTest,"a")
            img=cv2.imread(imgDataDir+x)
            img=img[80:399,160:479]
            img=cv2.resize(img,(227,227))
            cv2.imwrite(imgDataOutDir+x[0:8]+'_preprocessed_0.png',img)
            f.write(x[0:8]+'_preprocessed_0\n')
            f.close()
            bbsIn_all=np.loadtxt(posDataDir+"pcd{}cpos.txt".format(x[3:7]))
            m,n=np.shape(bbsIn_all)
            bbsnum=m//4
            bbsOut=np.zeros((m,n))
            lst=[]
            dev=0
            for idx in range(bbsnum):
                bbsin=bbsIn_all[idx*4:idx*4+4]
                # print(bbsin)

                 # print(bbsin)
                if np.isnan(bbsin).any():
                    dev=1
                    lst.extend([i for i in range(idx*4,idx*4+4)])
                    continue
                bbsinshift=bbsin-[320,240]
                # print(bbsinshift)
                bbsInShiftBack=(bbsinshift+[160,160])*cropSize/320
                bbsOut[idx*4:idx*4+4]=bbsInShiftBack
            if dev:
                bbsOut=np.delete(bbsOut,lst,axis=0)
            file_writeID=open(annotationDataOutDir+x[0:8]+"_preprocessed_0.txt","w")
            for j in range(len(bbsOut)//4):
                points=bbsOut[4*j:4*j+4]
                width=np.sqrt(sum(np.square(bbsOut[j*4]-bbsOut[j*4+1])))
                height=np.sqrt(sum(np.square(bbsOut[j*4+1]-bbsOut[j*4+2])))
                if bbsOut[j*4][0]>bbsOut[j*4+1][0]:
                    theta=math.atan((bbsOut[j*4+1][1]-bbsOut[j*4][1])/(bbsOut[j*4][0]-bbsOut[j*4+1][0]))
                else:
                    theta=math.atan((bbsOut[j*4][1]-bbsOut[j*4+1][1])/(bbsOut[j*4+1][0]-bbsOut[j*4][0]))
                x_crt,y_crt=sum(points)/4
                x_min=x_crt-width/2
                x_max=x_crt+width/2
                y_min=y_crt-height/2
                y_max=y_crt+height/2
                if x_min<0 or y_min<0 or x_max>227 or y_max>227:
                    continue
                # print(bbsOut[i])
                orient=round((theta/pi*180+90)/10)+1
                file_writeID.write('{} {} {} {} {}\n'.format(orient,x_min,y_min,x_max,y_max))
            file_writeID.close()

            continue
    
    img=cv2.imread(imgDataDir+x)
    bbsIn_all=np.loadtxt(posDataDir+"pcd{}cpos.txt".format(x[3:7]))
    imageout,bbsOut=dataPreprocessing_fasterrcnn(img,bbsIn_all,227,1,1)
    # print(imageout)
    for i in range(len(imageout)):
        # print(i)
        file_writeID=open(annotationDataOutDir+x[0:8]+"_preprocessed_{}.txt".format(i),"w")
        for j in range(len(bbsOut[i])//4):
            points=bbsOut[i][4*j:4*j+4]
            width=np.sqrt(sum(np.square(bbsOut[i][j*4]-bbsOut[i][j*4+1])))
            height=np.sqrt(sum(np.square(bbsOut[i][j*4+1]-bbsOut[i][j*4+2])))
            if bbsOut[i][j*4][0]>bbsOut[i][j*4+1][0]:
                theta=math.atan((bbsOut[i][j*4+1][1]-bbsOut[i][j*4][1])/(bbsOut[i][j*4][0]-bbsOut[i][j*4+1][0]))
            else:
                theta=math.atan((bbsOut[i][j*4][1]-bbsOut[i][j*4+1][1])/(bbsOut[i][j*4+1][0]-bbsOut[i][j*4][0]))
            x_crt,y_crt=sum(points)/4
            x_min=x_crt-width/2
            x_max=x_crt+width/2
            y_min=y_crt-height/2
            y_max=y_crt+height/2
            orient=round((theta/pi*180+90)/10)+1
            if x_min<0 or y_min<0 or x_max>227 or y_max>227:
                
                continue
            # print(bbsOut[i])
            
            file_writeID.write('{} {} {} {} {}\n'.format(orient,x_min,y_min,x_max,y_max))
        file_writeID.close()
        # print(imgDataOutDir+x[0:8]+"_preprocessed_{}.png".format(i))
        cv2.imwrite(imgDataOutDir+x[0:8]+"_preprocessed_{}.png".format(i),imageout[i])
        file_writeID=open(imgSetTrain,"a")
        file_writeID.write(x[0:8]+"_preprocessed_{}\n".format(i))
        file_writeID.close()

file_writeID=open(imgSetTrain1,"a")
for x in open(imgSetTrain):
    size = os.path.getsize('/home/sjy/grasp_data/data/Annotations/{}.txt'.format(x[:-1]))
    if size!=0:
        file_writeID.write(x[:-1]+'\n')
file_writeID.close()

        

其中输入地址和输出地址自己改一下就行。

注意，生成的数据类型的储存格式是

INRIA
|-- data
    |-- Annotations
         |-- *.txt (Annotation files)
    |-- Images
         |-- *.png (Image files)
    |-- ImageSets
         |-- train.txt

 这在后面会用到，准备好数据之后，主要需要改的地方是.lib/datasets/factory.py

43行改成自己的数据地址

graspRGB_devkit_path='/home/sjy/grasp_data'

之后运行

./experiments/scripts/train_faster_rcnn.sh 0 graspRGB res50

这里我遇到一个报错

TypeError: bottleneck() argument after ** must be a mapping, not tuple

解决方法，打开python3.6/site-packages/tensorflow/contrib/slim/python/slim/nets/resnet_utils.py

203行后加如下

unit={'depth':unit[0],'depth_bottleneck':unit[1],'stride':unit[2]}

折腾了这么久总算跑起来了

注意：重新训练需要删除data/cache 以及output