faster CPU inference, minor fixes

- in pointnet: decreased the batch size for CPU inference, now much faster
- in visualize: automatically look for feature_supervizion if features is not found
- in graph_processing: allows to compute spg on both train and test sets
- in README: incorporated @atineoSE fix for the recurring make error

README.md

@@ -70,15 +70,17 @@ cd build
 cmake .. -DPYTHON_LIBRARY=$CONDAENV/lib/libpython3.6m.so -DPYTHON_INCLUDE_DIR=$CONDAENV/include/python3.6m -DBOOST_INCLUDEDIR=$CONDAENV/include -DEIGEN3_INCLUDE_DIR=$CONDAENV/include/eigen3
 make
 ```
-The code was tested on Ubuntu 14.04 with Python 3.6 and PyTorch 0.2 to 1.0.
+The code was tested on Ubuntu 14 and 16 with Python 3.5 to 3.8 and PyTorch 0.2 to 1.1.
 
 ### Troubleshooting
 
-Common sources of error and how to fix them:
-- $CONDA_ENV is not well defined : define it or replace $CONDA_ENV by the absolute path of your environment (find it with ```locate anaconda```)
+Common sources of errors and how to fix them:
+- $CONDAENV is not well defined : define it or replace $CONDAENV by the absolute path of your conda environment (find it with ```locate anaconda```)
 - anaconda uses a different version of python than 3.6m : adapt it in the command. Find which version of python conda is using with ```locate anaconda3/lib/libpython```
 - you are using boost 1.62 or older: update it
 - cut pursuit did not download: manually clone it in the ```partition``` folder or add it as a submodule as proposed in the requirements, point 4.
+- error in make: `'numpy/ndarrayobject.h' file not found`: set symbolic link to python site-package with `sudo ln -s $CONDAENV/lib/python3.7/site-packages/numpy/core/include/numpy $CONDAENV/include/numpy`
+
 
 ## Running the code
 

learning/pointnet.py

@@ -209,7 +209,7 @@ def run_batch(self, model, clouds, clouds_global, *excess):
     def run_batch_cpu(self, model, clouds, clouds_global, *excess):
         """ Evaluates the cloud on CPU, but put the values in the CPU as soon as they are computed"""
         #cudnn cannot handle arrays larger than 2**16 in one go, uses batch
-        batch_size = 2**16-1
+        batch_size = 2**10-1
         n_batches = int(clouds.shape[0]/batch_size)
         emb_total = self.run_batch(model, clouds[:batch_size,:,:], clouds_global[:batch_size,:]).cpu()
         for i in range(1,n_batches+1):

partition/visualize.py

@@ -44,10 +44,9 @@
 if args.dataset == 'custom_dataset':
     n_labels = 10    
 #---load the values------------------------------------------------------------
-if args.supervized_partition:
+fea_file   = root + "features/"          + folder + file_name + '.h5'
+if not os.path.isfile(fea_file):
     fea_file   = root + "features_supervision/"          + folder + file_name + '.h5'
-else:
-    fea_file   = root + "features/"          + folder + file_name + '.h5'
 spg_file   = root + "superpoint_graphs/" + folder + file_name + '.h5'
 ply_folder = root + "clouds/"            + folder 
 ply_file   = ply_folder                  + file_name

supervized_partition/graph_processing.py

@@ -329,11 +329,12 @@ def create_sema3d_datasets(args, test_seed_offset=0):
         testlist += [path + 'train/' + f + '.h5' for f in train_names]
     if 'val' in args.db_test_name:
         testlist += [path + 'train/' + f + '.h5' for f in valid_names]
-    elif 'testred' in args.db_test_name:
+    if 'testred' in args.db_test_name:
         testlist += [f for f in glob.glob(path + 'test_reduced/*.h5')]
-    elif 'testfull' in args.db_test_name:
+    if 'testfull' in args.db_test_name:
         testlist += [f for f in glob.glob(path + 'test_full/*.h5')]
 
+
     return tnt.dataset.ListDataset(trainlist,
                                    functools.partial(graph_loader, train=True, args=args, db_path=args.ROOT_PATH)), \
            tnt.dataset.ListDataset(testlist,