Skip to content

Latest commit

 

History

History
40 lines (31 loc) · 1.93 KB

getting_started.md

File metadata and controls

40 lines (31 loc) · 1.93 KB
Raw

Getting Started

Data Preparation

Download the data (VOC, Cityscapes) and pre-trained models from OneDrive link:
DATA/
|-- city
|-- pascal_voc
|-- pytorch-weight
|   |-- resnet50_v1c.pth
|   |-- resnet101_v1c.pth

Training && Inference on PASCAL VOC:

$ cd ./model/voc8.res50v3+.CPS
$ bash script.sh
  • The tensorboard file is saved in log/tb/ directory.
  • In script.sh, you need to specify some variables, such as the path to your data dir, the path to your snapshot dir that stores checkpoints, etc.
  • We have released the training log and pretrained model for this experiment on OneDrive. The performance is slightly different (73.28) from that of paper (73.20) due to randomness.
  • We have also released the training log of city8.res50v3+.CPS.

Different Partitions

To try other data partitions beside 1/8, you just need to change two variables in config.py:

C.labeled_ratio = 8
C.nepochs = 34

Please note that, for fair comparison, we control the total iterations during training in each experiment similar (almost the same), including the supervised baseline and semi-supervised methods. Therefore, the nepochs for different partitions are different. We list the nepochs for different datasets and partitions in the below.

Dataset 1/16 1/8 1/4 1/2
VOC 32 34 40 60
Cityscapes 128 137 160 240