This Repo contains code and data for "RITA: Group Attention is All You Need for Timeseries Analytics."

• RITA
  • Installation
  • Data
  • Model Running
    • Full-label Training
    • Pretrain & Few-label Finetuning

In stall the required packages through pip install:

pip install -r requirements.txt

The zip file of preprocessed dataset can be acess at https://storage.googleapis.com/rita_resources/rita_dataset.tar.gz. Please download the zip file and put it under the directory rita/. Then perform the following commands to unzip dataset files.

tar -zxvf rita_dataset.tar.gz

Go to the under directory code:

cd code/

The running command is as following:

python run.py --dataset DATASET --mode MODE --data
_size LABLE_SIZE [--pretrained_path PRETRAINED_PATH]

DATASET describes the dataset, which should be among [wisdm,rwhar,hhar,ecg]. MODE describes the running mode, which should be among [pretrain,train,finetune]. LABLE_SIZE describes the size of labels used in training stage, which should be among [full,few]. PRETRAINED_PATH is requried when MODE=finetune; it indicates the path to pretrained checkpoint to start with.

To perform full-label training, set MODE to train and LABLE_SIZE to full. For example, the command of full-label training on dataset WISDM is:

python run.py --dataset wisdm --mode train --data_size full

To perform self-superviesed pretraining, set MODE to pretrain. For example, the command of full-label training on dataset WISDM is:

python run.py --dataset wisdm --mode pretrain --data_size full

The pretrained checkpoints are stored in directory rita/code/checkpoints. After pretraining, the checkpoints can be copied and saved for further finetuning.

To perform few-label finetuning, set MODE to finetune and set PRETRAINED_PATH to the pretrained checkpoint's path, which is either obtained by performing pretraining as described above or downloading the pretrained checkpoints at https://storage.googleapis.com/rita_resources/rita_pretrained.tar.gz. For example, the command of few-label finetuning on dataset WISDM is:

python run.py --dataset wisdm --mode finetune --data_size few --pretrained_path WISDM_PRETRAINED_PATH