update configs

config/softgroup_default_scannet.yaml

@@ -1,9 +1,3 @@
-GENERAL:
-  task: train  # train, test
-  manual_seed: 123
-  model_dir: model/softgroup/softgroup.py
-  dataset_dir: data/scannetv2_inst.py
-
 model:
   channels: 32
   num_blocks: 7
@@ -20,106 +14,59 @@ model:
                           8331., 3948., 3166., 5629., 11719.,
                           1003., 3317., 4912., 10221., 3889.,
                           4136., 2120., 945., 3967., 2589.]
+    npoint_thr: 0.05  # absolute if class_numpoint == -1, relative if class_numpoint != -1
+    ignore_classes: [0, 1]
   instance_voxel_cfg:
     scale: 50
     spatial_shape: 20
+  train_cfg:
+    max_proposal_num: 200
+    pos_iou_thr: 0.5
   test_cfg:
     x4_split: False
     cls_score_thr: 0.001
     mask_score_thr: -0.5
     min_npoint: 100
-  fixed_modules: []
+  fixed_modules: ['input_conv', 'unet', 'output_layer', 'semantic_linear', 'offset_linear']
 
 data:
   train:
+    type: 'scannetv2'
     data_root: 'dataset/scannetv2'
-    prefix: 'val'
+    prefix: 'train'
     suffix: '_inst_nostuff.pth'
+    training: True
     voxel_cfg:
       scale: 50
       spatial_shape: [128, 512]
       max_npoint: 250000
+      min_npoint: 5000
   test:
+    type: 'scannetv2'
     data_root: 'dataset/scannetv2'
     prefix: 'val'
     suffix: '_inst_nostuff.pth'
+    training: False
     voxel_cfg:
       scale: 50
       spatial_shape: [128, 512]
       max_npoint: 250000
-  data_loader:
+      min_npoint: 5000
+
+dataloader:
+  train:
     batch_size: 4
     num_workers: 4
-      
+  test:
+    batch_size: 1
+    num_workers: 16
 
-
-
-
-DATA:
-  data_root: dataset
-  dataset: scannetv2
-  filename_suffix: _inst_nostuff.pth
-  semantic_classes: 20
-  classes: 18
-  class_numpoint_mean: [-1., -1., 3917., 12056., 2303.,
-                        8331., 3948., 3166., 5629., 11719.,
-                        1003., 3317., 4912., 10221., 3889.,
-                        4136., 2120., 945., 3967., 2589.]
-  ignore_label: -100
-  input_channel: 3
-  scale: 50   # voxel_size = 1 / scale, scale 50 -> voxel_size 0.02m
-  batch_size: 4
-  full_scale: [128, 512]
-  max_npoint: 250000
-  mode: 4 # 4=mean
-
-STRUCTURE:
-  model_name: softgroup
-  width: 32
-  block_residual: True
-  block_reps: 2
-  use_coords: True
-  semantic_only: False
-
-train:
-  epochs: 500
-  train_workers: 4 # data loader workers
-  optim: Adam # Adam or SGD
+optimizer:
+  type: 'Adam'
   lr: 0.001
-  step_epoch: 200
-  multiplier: 0.5
-  momentum: 0.9
-  weight_decay: 0.0001
-  save_freq: 16  # also eval_freq
-  loss_weight: [1.0, 1.0, 1.0, 1.0, 1.0] # semantic_loss, offset_norm_loss, cls_loss, mask_loss, score_loss
-  fg_thresh: 1.
-  bg_thresh: 0.
-  score_scale: 50 # the minimal voxel size is 2cm
-  score_fullscale: 20  
-  score_mode: 4 # mean
-  pretrain_path: 'hais_ckpt.pth'
-  pretrain_module: ['input_conv', 'unet', 'output_layer', 'semantic_linear', 'offset_linear', 'intra_ins_unet', 'intra_ins_outputlayer']
-  fix_module: ['input_conv', 'unet', 'output_layer', 'semantic_linear', 'offset_linear']
-  point_aggr_radius: 0.04
-  cluster_shift_meanActive: 300
-  prepare_epochs: -1
-  max_proposal_num: 200
-  iou_thr: 0.5
-  score_thr: 0.2
 
-TEST:
-  split: val
-  test_epoch: 500
-  test_workers: 16
-  test_seed: 567
-  using_NMS: False
-  TEST_NMS_THRESH: 0.3
-  TEST_SCORE_THRESH: -1
-  TEST_NPOINT_THRESH: 100
-
-  eval: True
-  save_semantic: False
-  save_pt_offsets: False
-  save_instance: False
-
-  test_mask_score_thre: -0.5 # bias fg << bg
+epochs: 512
+step_epoch: 200
+save_freq: 8
+pretrain: 'hais_ckpt.pth'
+work_dir: 'work_dirs/softgroup_scannet'

config/softgroup_fold5_backbone_s3dis.yaml

@@ -12,9 +12,14 @@ model:
     mean_active: 300
     class_numpoint_mean: [1823, 7457, 6189, 7424, 34229, 1724, 5439,
                           6016, 39796, 5279, 5092, 12210, 10225]
+    npoint_thr: 0.05  # absolute if class_numpoint == -1, relative if class_numpoint != -1
+    ignore_classes: [0, 1]
   instance_voxel_cfg:
     scale: 50
     spatial_shape: 20
+  train_cfg:
+    max_proposal_num: 200
+    pos_iou_thr: 0.5
   test_cfg:
     x4_split: True
     cls_score_thr: 0.001
@@ -46,9 +51,14 @@ data:
       spatial_shape: [128, 512]
       max_npoint: 250000
       min_npoint: 5000
-  dataloader:
+
+dataloader:
+  train:
     batch_size: 4
     num_workers: 4
+  test:
+    batch_size: 1
+    num_workers: 1
 
 optimizer:
   type: 'Adam'

config/softgroup_fold5_default_s3dis.yaml

@@ -51,9 +51,14 @@ data:
       spatial_shape: [128, 512]
       max_npoint: 250000
       min_npoint: 5000
-  dataloader:
+
+dataloader:
+  train:
     batch_size: 4
     num_workers: 4
+  test:
+    batch_size: 1
+    num_workers: 1
 
 optimizer:
   type: 'Adam'

data/__init__.py

@@ -29,11 +29,11 @@ def build_dataloader(dataset, batch_size=1, num_workers=1, training=True):
             drop_last=True,
             pin_memory=True)
     else:
-        assert batch_size == 1 and num_workers == 1
+        assert batch_size == 1
         return DataLoader(
             dataset,
-            batch_size=1,
-            num_workers=1,
+            batch_size=batch_size,
+            num_workers=num_workers,
             collate_fn=dataset.collate_fn,
             shuffle=False,
             drop_last=False,

data/scannetv2.py

@@ -8,7 +8,7 @@ class ScanNetDataset(CustomDataset):
                'bathtub', 'otherfurniture')
 
     def getInstanceInfo(self, xyz, instance_label, label):
-        ret = super().getInstanceInfo(xyz, instnace_label, label)
+        ret = super().getInstanceInfo(xyz, instance_label, label)
         instance_num, instance_pointnum, instance_cls, pt_offset_label = ret
         instance_cls = [x - 2 if x != -100 else x for x in instance_cls]
         return instance_num, instance_pointnum, instance_cls, pt_offset_label

test.py

@@ -82,7 +82,7 @@ if __name__ == '__main__':
     model.cuda()
 
     dataset = build_dataset(cfg.data.test, logger)
-    dataloader = build_dataloader(dataset, training=False)
+    dataloader = build_dataloader(dataset, training=False, **cfg.dataloader.test)
     all_preds, all_gts = [], []
     with torch.no_grad():
         model = model.eval()

train.py

@@ -87,8 +87,8 @@ if __name__ == '__main__':
     # data
     train_set = build_dataset(cfg.data.train, logger)
     val_set = build_dataset(cfg.data.test, logger)
-    train_loader = build_dataloader(train_set, training=True, **cfg.data.dataloader)
-    val_loader = build_dataloader(val_set, training=False)
+    train_loader = build_dataloader(train_set, training=True, **cfg.dataloader.train)
+    val_loader = build_dataloader(val_set, training=False, **cfg.dataloader.test)
 
     # optim
     optimizer = build_optimizer(model, cfg.optimizer)