update

README.md

@@ -20,21 +20,22 @@ root
 ## environment
 1. pytorch
 refer to OctAttention-obj-pytorch/README.md
+
 2. tensorflow
-tensorflow-gpu 2.3.1
-others same as pytorch
+refer to requirements.txt
 
 ## command
-* pytorch && tensorflow:
+* tensorflow:
+> cd TensorFlow
 
 training:
 >python octAttention.py 
 
 encode:
->python encoder.py  
+>python encoder.py --input="/userhome/PCGCv1/pytorch_eval/28_airplane_0270.ply" --ckpt_dir="checkpoint_TF_1024"
 
 decode:
->python decoder.py 
+>python decoder.py --input="/userhome/PCGCv1/pytorch_eval/28_airplane_0270.ply"  --ckpt_dir="checkpoint_TF_1024"
 
 ## performance
 * test on some PC files. From the result below, we can see that, encoding in pytorch is faster than tensorflow, maybe because of the fast operation of linear module of pytorch. And bpip of tensorflow is closed to pytorch.

TensorFlow/decoder.py

@@ -121,7 +121,7 @@ if __name__=="__main__":
 
     for oriFile in list_orifile: # from encoder.py
         ptName = os.path.basename(oriFile)[:-4]
-        matName = '../Data/testPly/'+ptName+'.mat'
+        matName = './Data/testPly/'+ptName+'.mat'
         binfile = expName+'/data/'+ptName+'.bin'
         cell,mat =matloader(matName)
 
@@ -141,4 +141,4 @@ if __name__=="__main__":
         # Dequantization
         DQpt = (ptrec*qs+offset)
         pt.write_ply_data(expName+"/temp/test/rec.ply",DQpt)
-        pt.pcerror(p,DQpt,None,'-r 1',None).wait()
+        # pt.pcerror(p,DQpt,None,'-r 1',None).wait()

TensorFlow/encoder.py

@@ -12,19 +12,30 @@ from networkTool import CPrintl,expName
 from octAttention import model
 import glob,datetime,os
 import pt as pointCloud #无框架
+import argparse
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument("--input", default='', help="Input filename.") #'compressed/28_airplane_0270' '/userhome/PCGCv1/testdata/8iVFB/redandblack_vox10_1550.ply'
+    parser.add_argument("--ckpt_dir", type=str, default='/userhome/PCGCv1/pytorch2/ckpts/hyper_mgpu2/epoch_13_12599.pth', dest="ckpt_dir",  help='checkpoint')
+    args = parser.parse_args()
+    return args
+
 ############## warning ###############
 ## decoder.py relys on this model here
 ## do not move this lines to somewhere else
 # model = model.to(device)
 # saveDic = reload(None,'modelsave/obj/encoder_epoch_00800093.pth')
 # model.load_state_dict(saveDic['encoder'])
+args = parse_args()
 input_data = tf.ones([1024, 32, 4, 6],dtype=tf.int32)
 input_mask = model.generate_square_subsequent_mask(1024)
 output = model(input_data, input_mask,[],training=False)
-model.load_weights(tf.train.latest_checkpoint('/userhome/OctAttention/OctAttention-obj/Exp/Obj/checkpoint_TF/')) #文件夹名称
+model.load_weights(tf.train.latest_checkpoint(args.ckpt_dir)) #文件夹名称
 
 ###########Objct##############
-list_orifile = ['../28_airplane_0270.ply']
+list_orifile = [args.input]
 if __name__=="__main__":
     printl = CPrintl(expName+'/encoderPLY_tf.txt') #输出的同时保存到log
     printl('_'*50,'OctAttention V tf','_'*50)
@@ -38,8 +49,8 @@ if __name__=="__main__":
         ptName = os.path.splitext(os.path.basename(oriFile))[0] 
         for qs in [1]:
             ptNamePrefix = ptName
-            matFile,DQpt,refPt = dataPrepare(oriFile,saveMatDir='../Data/testPly',qs=qs,ptNamePrefix='',rotation=False)
+            matFile,DQpt,refPt = dataPrepare(oriFile,saveMatDir='./Data/testPly',qs=qs,ptNamePrefix='',rotation=False)
             # please set `rotation=True` in the `dataPrepare` function when processing MVUB data
             main(matFile,model,actualcode=True,printl =printl) # actualcode=False: bin file will not be generated
-            print('_'*50,'pc_error','_'*50)
-            pointCloud.pcerror(refPt,DQpt,None,'-r 1023',None).wait() #这里比的是量化前，量化后再反量化的点
+            # print('_'*50,'pc_error','_'*50)
+            # pointCloud.pcerror(refPt,DQpt,None,'-r 1023',None).wait() #这里比的是量化前，量化后再反量化的点

TensorFlow/networkTool.py

@@ -15,8 +15,8 @@ os.environ["CUDA_VISIBLE_DEVICES"] = "0"
 
 # Network parameters
 bptt = 1024 # Context window length
-expName = '../Exp/Obj'
-DataRoot = '../Data/Obj'
+expName = './Exp/Obj'
+DataRoot = './Data/Obj'
 
 checkpointPath = expName+'/checkpoint_TF'
 levelNumK = 4

requirements.txt

@@ -0,0 +1,8 @@
+h5py==2.10.0
+hdf5storage==0.1.18
+numpy==1.18.4
+pandas==1.0.4
+Pillow==9.1.0
+plyfile==0.7.4
+tensorflow_gpu==2.3.1
+tqdm==4.48.0