IKKIM00 / temporal-fusion-contrastive-learning

TFCL (Temporal Fusion Contrastive Learning)

This repository is the official implementation of "Human Activity Recognition via Temporal Fusion Contrastive Learning" [paper].

Model Architecture

Requirmentss

• python == 3.9
• pytorch == 1.10.1
• pytorch-lightning == 1.6.4

How to start

1. Preprocess dataset

A. Download dataset from following urls

• MobiAct Dataset
• DRL Dataset

B. Run following ipynb files for each dataset

• MobiAct Datset - datasets/preprocess_mobiact.ipynb
• DLR Dataset - datasets/preprocess_dlr.ipynb

2. Hyper-parameters for main.py

• --experiment description : set experiment description. Default : Exp1
• --seed: seed value. Default: 42
• --model_type : set model. Default : TFCL, for baseline models: SimclrHAR, CSSHAR, CPCHAR
• --training_mode: choose training mode between self_supervised, fine_tune, train_linear, supervised. Default: supervised
• --loss_func: choose between focal and cross_entropy. Default: cross_entropy
• --batch_size: set batch_size. Default: 512
• --aug_method1: set first augmentation method. Default: jitter_scale
• --aug_method2: set second augmentation method. Default: permutation_jitter
• --static_use: choose whether to use static data. Use static data: --static_use, not use static data: --no-static_use
• --sampler_use: choose whether to use imbalance dataset sampler. Use sampler: --sampler_use, not use sampler use: --no-sampler_use
• --dataset: choose dataset. Default: mobiact
• --logs_save_dir: saving directory. Default: experiments_logs
• --device: choose device. Default: cpu
• --home_path: home directory. Default: current directory

For aug_method1 and aug_method2, choose between ["jitter" "scale" "jitter_scale" "permutation" "permutation_jitter" "rotation" "invert" "timeflip" "shuffle" "warp"]

3. Training

with static data + sampler

Terminal command

python main.py --model_type 'model_type' --sampler_use --dataset 'dataset_name' --device 'devce' --training_mode 'train_mode' --loss_func 'loss_function' --aug_method1 'first_augmentation_method' --aug_method2 'second_augmentation_method'

Example

python main.py --model_type TFCL --sampler_use --dataset mobiact --device 0,1 --training_mode self_supervised --loss_func focal --aug_method1 jitter_scale --aug_method2 permutation_jitter

with static data + no sampler

Terminal command

python main.py --model_type 'model_type' --no-sampler_use --dataset 'dataset_name' --device 'devce' --training_mode 'train_mode' --loss_func 'loss_function' --aug_method1 'first_augmentation_method' --aug_method2 'second_augmentation_method'

Example

python main.py --model_type TFCL --no-sampler_use --dataset mobiact --device 0,1 --training_mode self_supervised --loss_func focal --aug_method1 jitter_scale --aug_method2 permutation_jitter

without static data + sampler

Terminal command

python main.py --model_type 'model_type' --sampler_use --no-static_use --dataset 'dataset_name' --device 'devce' --training_mode 'train_mode' --loss_func 'loss_function' --aug_method1 'first_augmentation_method' --aug_metho2 'second_augmentation_method'

Example

python main.py --model_type TFCL --sampler_use --no-static_use --dataset mobiact --device 0,1 --training_mode self_supervised --loss_func focal --aug_metho1 jitter_scale --aug_method2 permutation_jitter

without static data + no sampler

Terminal command

python main.py --model_type 'model_type' --no-sampler_use --no-static_use --dataset 'dataset_name' --device 'devce' --training_mode 'train_mode' --loss_func 'loss_function' --aug_method1 'first_augmentation_method' --aug_method2 'second_augmentation_method'

Example

python main.py --model_type TFCL --no-sampler_use --no-static_use --dataset mobiact --device 0,1 --training_mode self_supervised --loss_func focal --aug_method1 jitter_scale --aug_method2 permutation_jitter