Skip to content

Latest commit

 

History

History
58 lines (48 loc) · 2.3 KB

Training_Guidelines.md

File metadata and controls

58 lines (48 loc) · 2.3 KB
Raw

Training Guideline

1. Overview of MIST framework

Structure of MIST The framework consists of two stages, i.e., Pseudo labels generation and Feature Encoder Finetuning. For the first stage, we train a MIL-based pseudo labels generator with pre-extracted features. As for the other stage, we train the self-guided attention enhanced feature encoder with the pseudo labels.

2. Environment Preparation

  • python>=3.6
  • apex
  • pytorch=1.5.0+cu101
  • torchvision=0.6.0+cu101
  • tensorboardX
  • h5py
  • opencv
  • scikit-learn
  • yacs

[ Attention!! ] Before doing any further, remember that the paths in the codes may need to be modified to adapt to your environment.

3. Data Preparation

Pre-extracted features for ShanghaiTech are uploaded on SHT_Feats_for_MIL. Moreover the Kinetics pretrained I3D model and Sport-100M pretrained C3D model are uploaded on pretrained_models. You should download all of them and place them in the proper place as configs/constant.py indicates.

Specifically, the test_frame_mask of ShanghaiTech is downloaded from Download. test_frame_mask is uploaded in data/test_frame_mask/ now.

[Update!] As the original ShanghaiTech dataset link is not worked now. I uploaded the h5py file for ShanghaiTech and its corresponing annotations are uploaded on [BaiduYun] with multiple sub-files, you can open/unzip it with WinRAR

BaiduYun link, code：kym5

4. Stage 1 training

The stage 1 is to generate pseudo labels. First we train the MIL-based generator.

python stage1training/train_MIL_generator.py

Then, generate pseudo labels with the command below:

python stage1training/train_MIL_generator.py --generate_PL

5. Stage 2 training

You are recommend to use the pseudo labels generated by us, which is placed in data/.

The commands of training C3D/I3D for ShanghaiTech and UCF-Crime dataset are writen in bash files. You can execute the command as below:

cd stage2training
% for SHT_C3D
bash SHT_C3D_train.sh
% for SHT_I3D
bash SHT_I3D_train.sh
% for UCF_C3D
bash UCF_C3D_train.sh