Skip to content

Latest commit

 

History

History

sem_seg_dense

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 

Sem_seg_dense and sem_seg_sparse are both for the semantic segmentation task. The difference between them is that the data shape for the graph is different. As for sem_seg_sparse, data shape is N x C (N is number of nodes, C is feature_size) and there is a batch variable indicating the batch of each node. But for sem_seg_dense, data shape is B x C x N x 1 (B is batch size, N here means the number of nodes per graph).
In gcn_lib, there are two folders: dense and sparse. They are used for different data shapes above.

Train

We keep using 2 Tesla V100 GPUs for distributed training.

cd examples/sem_seg/dense

Train ResGCN-28 (DeepGCN with 28 residually connected EdgeConv layers, dilated graph convolutions and batch normalization):

CUDA_VISIBLE_DEVICES=0,1 python train.py --multi_gpus --phase train --data_dir /data/deepgcn/S3DIS

Just need to set --data_dir $yourpath/to/data, dataset will be downloaded automatically.

If you want to train model with other gcn layers (for example, mrgcn), run

python train.py --conv mr --multi_gpus --phase train --data_dir /data/deepgcn/S3DIS 

Other parameters for changing the architecture are:

--block         graph backbone block type {res, plain, dense}
--conv          graph conv layer {edge, mr}
--n_filters     number of channels of deep features, default is 64
--n_blocks      number of basic blocks, default is 28

A shallow version of DeepGCN (ResGCN-7) could be trained by the command below:

CUDA_VISIBLE_DEVICES=0 python train.py  --multi_gpus --phase train --data_dir /data/deepgcn/S3DIS --n_blocks 7

Evaluation

Quick test on area 5, run:

python test.py --pretrained_model checkpoints/sem_seg_dense-res-edge-28-64-ckpt_best_model.pth  --batch_size 32  --data_dir /data/deepgcn/S3DIS 

Pretrained Models

Our pretrained model is available here google driver.

Note: Please use our Tensorflow code if you want to reproduce the same result in the paper. The performance of pytorch code is slightly worse than tensorflow. mIOU is 52.11% on Area 5 compared to 52.49% in the tensorflow version.

python test.py --pretrained_model checkpoints/sem_seg_dense-res-edge-28-64-ckpt_best_model.pth  --batch_size 32  --data_dir /data/deepgcn/S3DIS

Lower the batch size if running out of memory. The batch size will not influence the test results.