Skip to content

Latest commit

 

History

History
50 lines (33 loc) · 2.44 KB

README.md

File metadata and controls

50 lines (33 loc) · 2.44 KB
Raw

Pose-based Action Recognition: Robust Models with Contrastive Loss

Semester project

Student: Aleksandra Novikova, Data Science

Supervisor: Mohamed Ossama Ahmed Abdelfattah

Navigating the repo

  • Report.pdf - project report

  • TCL/ - implementation of the contrastive loss method with noisy data for training robust models. The main code is taken from the repository TCL. Models (backbones) are taken from repositories MotionBERT and pySKL.

  • pyskl/ - implementation of different input types for a PoseConv3D model. The main code is taken from the repository pySKL.

Details can be found in the respective repositories.

Training

Contrastive learning

  • PoseConv3D model training:
python main.py ntu60 RGB --seed 123 --input_f Skeleton --in_channels 68 --strategy classwise --arch resnet18 --num_segments 8 --second_segments 8 --threshold 0.8 --gd 20  --epochs 1000 --percentage 0.95 -j 2 --dropout 0.5 --consensus_type=avg --eval-freq=1 --print-freq 50 --shift --shift_div=8 --shift_place=blockres --npb --gpus 0  --mu 1 --gamma 5 --gamma2 0 --use_group_contrastive --sup_thresh 0 --batch-size 32 --valbatchsize 32  --lr_backbone1 0.001 --lr_decay1 0.9  --noise_alpha 0.1
  • MotionBERT model training:
python main.py ntu60 RGB --seed 123 --input_f Skeleton --in_channels 68 --strategy classwise --arch resnet18 --num_segments 8 --second_segments 8 --threshold 0.8 --gd 20  --epochs 1000 --percentage 0.95 -j 2 --dropout 0.5 --consensus_type=avg --eval-freq=1 --print-freq 50 --shift --shift_div=8 --shift_place=blockres --npb --gpus 0  --mu 1 --gamma 5 --gamma2 0 --use_group_contrastive --sup_thresh 0 --batch-size 32 --valbatchsize 32  --lr_backbone1 0.001 --lr_decay1 0.9  --noise_alpha 0.1 --model_type motionbert

where

  • noise_alpha - percentage of added noise to data
  • gamma - coefficient before instance contrastive loss
  • gamma2 - coefficient before group contrastive loss

PoseConv3D with different input types

bash tools/dist_train.sh configs/posec3d/slowonly_r50_ntu60_xsub/joint.py 1 --validate --test-last --test-best
  • configs/posec3d/slowonly_r50_ntu60_xsub/joint.py for 3D Heatmaps
  • configs/posec3d/slowonly_r50_ntu60_xsub/joint_grayscale.py for Grayscale
  • configs/posec3d/slowonly_r50_ntu60_xsub/joint_skeleton.py for Skeleton