QLab2025

A collection of two research-oriented projects focused on Time Series analysis and Image Restoration.

Overview

• Time Series:

  • P1T1: Multivariate Time Series Imputation using Transformer Encoder
  • P1T2: Time Series Forecasting using DLinear
  • P1T3: Multivariate Time Series Forecasting using Conditional Diffusion Model (DDPM)

• Image Restoration:

  • P2T1: Denoising using DNCNN Architecture
  • P2T2: Denoising using Unet Architecture

Time Series

Datasets

• Exchange Dataset: Used for P1T1 and P1T2. Download
• Timeseries Train Dataset: Used for P1T3. Download
• Timeseries Test Dataset: Used for P1T3. Download

Tasks

Each task is associated with a specific filename within the project repository.

P1T1: Multivariate Time Series Imputation using Transformer Encoder

• Filename: P1T1
• Objective: Train a transformer encoder to impute missing values in multivariate time series data with mask ratios of 12.5%, 25%, 37.5%, and 50%.
• Sequence Length: 96
• Evaluation Metrics: MSE, MAE
• Visualization: Plot results for different mask ratios.
• Resources: Use any online resources for model development.

P1T2: Time Series Forecasting using DLinear

• Filename: P1T2
• Objective: Implement and train the DLinear model for time series forecasting.
• Sequence Length: 96
• Prediction Length: 14
• Settings: Univariate (last feature) and Multivariate
• Evaluation Metrics: MSE, MAE
• Visualization: Compare results for univariate and multivariate settings.
• Reference: DLinear Paper

P1T3: Multivariate Time Series Forecasting using Conditional Diffusion Model (DDPM)

• Filename: P1T3
• Objective: Develop a conditional diffusion model for multivariate time series forecasting.
• Sequence Length: 96
• Prediction Length: 14
• Settings: Univariate (last feature) and Multivariate
• Evaluation Metrics: MSE, MAE
• Visualization: Compare results for univariate and multivariate settings.

Experiment Tracking

• Weights & Biases (W&B) Dashboard: Track and visualize experiments. Dashboard URL

---

Image Restoration

Dataset

• CBSD68 Dataset: Used for both P2T1 and P2T2. Clone from GitHub Repository
  • Use noisy35 and original folders as paired sets for training denoiser networks.

Tasks

Each task is associated with a specific filename within the project repository.

P2T1: Denoising using DNCNN Architecture

• Filename: P2T1
• Objective: Train a denoiser network using the DNCNN architecture.
• Evaluation Metrics: PSNR, SSIM
• Visualization: Display denoised images and metrics.
• Reference: DNCNN Paper

P2T2: Denoising using Unet Architecture

• Filename: P2T2
• Objective: Train a denoiser network using the Unet architecture.
• Evaluation Metrics: PSNR, SSIM
• Visualization: Display denoised images and metrics.

References

• Zhou, Y. et al., "DnCNN: Beyond a Gaussian Denoiser" (2017).
• Zeng, F. et al., "DLinear: Simplifying Time Series Forecasting with Linear Models" (2022).
• Ho, J. et al., "Denoising Diffusion Probabilistic Models" (2020).

---