Pytorch PVCNN:Point-Voxel CNN for Efficient 3D Deep Learning #3

1. データのダウンロード
下記、Shell を MSYS2 で実行して、ダウンロード、解凍を行います。
download_data.sh
#!/bin/bash # Download HDF5 for indoor 3d semantic segmentation (around 1.6GB) wget --no-check-certificate https://shapenet.cs.stanford.edu/media/indoor3d_sem_seg_hdf5_data.zip unzip indoor3d_sem_seg_hdf5_data.zip #rm indoor3d_sem_seg_hdf5_data.zip

2. s3dis.py を改造します。
datasets/s3dis_nishi.py
import os import os.path as osp import torch import h5py import numpy as np from torch.utils.data import Dataset import sys __all__ = ['S3DIS'] ORIGIN=False class _S3DISDataset_Access(): url = ('https://shapenet.cs.stanford.edu/media/' 'indoor3d_sem_seg_hdf5_data.zip') def __init__(self, root, test_area=6, train=True): assert test_area >= 1 and test_area <= 6 self.test_area = test_area self.root=root #super(S3DIS, self).__init__(root, transform, pre_transform, pre_filter) #path = self.processed_paths[0] if train else self.processed_paths[1] #self.data, self.slices = torch.load(path) self.name = self.url.split('/')[-1].split('.')[0] #print('self.name=',self.name) self.raw_dir=self.root+'/'+self.name #print('self.raw_dir=',self.raw_dir) self.raw_paths=[self.raw_dir+'/'+self.raw_file_names[0],self.raw_dir+'/'+self.raw_file_names[1]] @property def raw_file_names(self): return ['all_files.txt', 'room_filelist.txt'] def get_rooms(self): with open(self.raw_paths[1], 'r') as f: rooms = f.read().split('\n')[:-1] return rooms def get_all_data(self): print('self.raw_paths', self.raw_paths) with open(self.raw_paths[0], 'r') as f: filenames = [x.split('/')[-1] for x in f.read().split('\n')[:-1]] with open(self.raw_paths[1], 'r') as f: rooms = f.read().split('\n')[:-1] xs, ys = [], [] for filename in filenames: print('filename=',filename) f = h5py.File(osp.join(self.raw_dir, filename)) num_windows = f['data'].shape[0] # test by nishi if False: print('f[\'data\'].shape=',f['data'].shape) print('num_windows=',num_windows) sys.exit() xs += torch.from_numpy(f['data'][:]).unbind(0) ys += torch.from_numpy(f['label'][:]).to(torch.long).unbind(0) test_area = 'Area_{}'.format(self.test_area) self.train_data_list, self.test_data_list = [], [] for i, (x, y,r) in enumerate(zip(xs, ys ,rooms)): if False: #x.shape= torch.Size([4096, 9]) #y.shape= torch.Size([4096]) print('x[0]=',x[0]) print('y[0]=',y[0]) visual_o3dx3(x,trans=False,fs=9,norm=False) sys.exit() data = {'x':x,'y':y,'r':r} if test_area not in rooms[i]: #print('x.shpae=',x.shape,'y.shape=',y.shape,'r=',r) self.train_data_list.append(data) else: #print('x.shpae=',x.shape,'y.shape=',y.shape,'r=',r) self.test_data_list.append(data) print('len(self.train_data_list)',len(self.train_data_list)) print('len(self.test_data_list)=',len(self.test_data_list)) def get_lng(self,split): if split=='train': return len(self.train_data_list) else: return len(self.test_data_list) def get_data(self,split): if split=='train': return self.train_data_list else: return self.test_data_list class _S3DISDataset(Dataset): def __init__(self, root, num_points, data_list=None,data_lng=None,split='train', with_normalized_coords=True, holdout_area=5): """ :param root: directory path to the s3dis dataset :param num_points: number of points to process for each scene :param split: 'train' or 'test' :param with_normalized_coords: whether include the normalized coords in features (default: True) :param holdout_area: which area to holdout (default: 5) """ assert split in ['train', 'test'] self.root = root self.split = split self.num_points = num_points self.data_list=data_list self.data_lng=data_lng self.holdout_area = None if holdout_area is None else int(holdout_area) self.with_normalized_coords = with_normalized_coords # keep at most 20/30 files in memory self.cache_size = 20 if split == 'train' else 30 self.cache = {} # mapping batch index to corresponding file areas = [] # These informations are in room_filelist.txt by nishi if self.split == 'train': for a in range(1, 7): if a != self.holdout_area: #areas.append(os.path.join(self.root, f'Area_{a}')) areas.append(f'Area_{a}') else: areas.append(f'Area_{self.holdout_area}') print('areas=',areas) self.num_scene_windows, self.max_num_points = 0, 0 index_to_filename, scene_list = [], {} filename_to_start_index = {} # comment by nishi # get data from zero_0.h5 half_0.h5 if ORIGIN: for area in areas: # room_filelist.txt area_scenes = os.listdir(area) area_scenes.sort() for scene in area_scenes: current_scene = os.path.join(area, scene) scene_list[current_scene] = [] for split in ['zero', 'half']: current_file = os.path.join(current_scene, f'{split}_0.h5') filename_to_start_index[current_file] = self.num_scene_windows h5f = h5py.File(current_file, 'r') num_windows = h5f['data'].shape[0] self.num_scene_windows += num_windows for i in range(num_windows): index_to_filename.append(current_file) scene_list[current_scene].append(current_file) else: print('test') #sys.exit() self.index_to_filename = index_to_filename self.filename_to_start_index = filename_to_start_index self.scene_list = scene_list def __len__(self): #return self.num_scene_windows return self.data_lng def __getitem_org__(self, index): filename = self.index_to_filename[index] if filename not in self.cache.keys(): h5f = h5py.File(filename, 'r') scene_data = h5f['data'] scene_label = h5f['label_seg'] scene_num_points = h5f['data_num'] if len(self.cache.keys()) < self.cache_size: self.cache[filename] = (scene_data, scene_label, scene_num_points) else: victim_idx = np.random.randint(0, self.cache_size) cache_keys = list(self.cache.keys()) cache_keys.sort() self.cache.pop(cache_keys[victim_idx]) self.cache[filename] = (scene_data, scene_label, scene_num_points) else: scene_data, scene_label, scene_num_points = self.cache[filename] internal_pos = index - self.filename_to_start_index[filename] current_window_data = np.array(scene_data[internal_pos]).astype(np.float32) current_window_label = np.array(scene_label[internal_pos]).astype(np.int64) current_window_num_points = scene_num_points[internal_pos] choices = np.random.choice(current_window_num_points, self.num_points, replace=(current_window_num_points < self.num_points)) data = current_window_data[choices, ...].transpose() label = current_window_label[choices] # data[9, num_points] = [x_in_block, y_in_block, z_in_block, r, g, b, x / x_room, y / y_room, z / z_room] if self.with_normalized_coords: return data, label else: return data[:-3, :], label def __getitem__(self, index): #print('__getitem__()') dt = self.data_list[index] current_window_data=dt['x'].cpu().numpy() current_window_label=dt['y'].cpu().numpy() if False: print('type(current_window_data)=',type(current_window_data)) print('current_window_data.shape=',current_window_data.shape) print('type(current_window_label)=',type(current_window_label)) print('current_window_label.shape=',current_window_label.shape) current_window_num_points=current_window_data.shape[0] choices = np.random.choice(current_window_num_points, self.num_points, replace=(current_window_num_points < self.num_points)) data = current_window_data[choices, ...].transpose() label = current_window_label[choices] # data[9, num_points] = [x_in_block, y_in_block, z_in_block, r, g, b, x / x_room, y / y_room, z / z_room] if self.with_normalized_coords: return data, label else: return data[:-3, :], label class S3DIS(dict): def __init__(self, root, num_points, split=None, with_normalized_coords=True, holdout_area=5,data_list=None): super().__init__() s3dis_acc=_S3DISDataset_Access(root,test_area=holdout_area) #s3dis.download() s3dis_acc.get_all_data() if split is None: split = ['train', 'test'] elif not isinstance(split, (list, tuple)): split = [split] for s in split: data_list=s3dis_acc.get_data(s) data_lng=s3dis_acc.get_lng(s) self[s] = _S3DISDataset(root=root, num_points=num_points, split=s,data_list=data_list,data_lng=data_lng, with_normalized_coords=with_normalized_coords, holdout_area=holdout_area) if __name__ == '__main__': #root='E:/tmp/pvcnn/data/s3dis' root='E:/tmp/pointnet/sem_seg' s3dis=S3DIS(root=root,num_points=4098,holdout_area=6) index=0 (dt,label)=s3dis['train'][index] print('dt.shape=',dt.shape) print('label.shape=',label.shape) print('len(s3dis)=',len(s3dis['train'])) (dt,label)=s3dis['test'][index] print('dt.shape=',dt.shape) print('label.shape=',label.shape) print('len(s3dis)=',len(s3dis['test']))

3. configs/s3dis/__init__.py を修正。

下記例では、download_data.sh を、
E:/tmp/pointnet/sem_seg で実行した場合を想定しています。
E:/tmp/pointnet/sem_seg/indoor3d_sem_seg_hdf5_data ディレクトリーが、
解凍した、ディレクトリーです。

... #from datasets.s3dis import S3DIS # changed by nishi from datasets.s3dis_nishi import S3DIS ... # dataset configs configs.dataset = Config(S3DIS) #configs.dataset.root = 'data/s3dis/pointcnn' # changed by nishi configs.dataset.root = 'E:/tmp/pointnet/sem_seg' # zip を解凍したパス .... # train configs configs.train = Config() #configs.train.num_epochs = 50 # changed by nishi configs.train.num_epochs = 2 # ちょっとずつ増やす。 ....

4. 学習の実行。
> python train.py configs/s3dis/pvcnn/area5/c1.py --devices 0

1回、学習が終了したら、configs.train.num_epochs = XX を増やして、
何回かに分けて学習を行います。

5. Tensorboard の確認
> tensorboard --logdir=runs\s3dis.pvcnn.area5.c1

どうぞね、旨く動くかね?