ENCODiT

Error flagging and Neutralization using Conformal Out-of-Distribution Detection in Time-Series Data

CS637A Course Project

November, 2023

Group 14

1. Kartik Anant Kulkarni (210493)
2. Rishi Agarwal (210849)
3. Emaad Ahmed (210369)
4. Dhruva Singh Sachan (210343)

How to Run

• Download the dataset and the trained models from here.
• Add the dataset to ./dataset folder as per the organisation specified in ./dataset/Dataset.md.
• Based on the module to test go to the necessary sections.
• cd to the root directory.
• Create a new Conda Environment ENCODiT from the environment.yml and activate it.
• Pre-process the dataset with % python ./dataset.py. Corresponding output will be saved to the dataset directory.

Dynamic Window Module

• Execute: % cd ./dynamic_window. All future commands should run from this directory only.
• If the dataset has not been processed already, run % python src/preprocess.py. The video clips will be processed and saved to the dataset folder.
• Create dynamic_window/log and dynamic_window/dump folders to save the model and the fisher values.
• Run, % python src/train.py by setting appropriate window size and dimension of layer 4 of the LeNet Model. The trained model will be saved to the dynamic_window/log folder.
• From main repository directory run, % python src/inference.py by setting appropriate window size and choosing model.

RNN

• Execute: % cd ./rnn. All future commands should run from this directory only.
• If the dataset has not been processed already, run % python src/data.py. The video clips will be processed and saved to the dataset folder.
• Run, % python src/train.py by setting necessary model parameters and paths, for training the model.
• Update the model checkpoint path in inference.py to the newly trained model path. From main repository directory run, % python src/inference.py by selecting the appropriate model.