DeepResAutoencoder: Spectral Data Reconstruction using Residual Autoencoders

This project implements a deep residual autoencoder for unsupervised reconstruction of spectral data using PyTorch. The model is optimized using a composite loss that combines MSE, cosine similarity, and spectral smoothness. Visual diagnostics include spectral plots and UMAP latent space projection.

---

🚀 Features

• Residual skip connections for better learning
• Composite loss (MSE + Cosine + Smoothness)
• Mixed-Precision training with AMP (if CUDA available)
• Early stopping & OneCycleLR scheduler
• Spectral reconstruction and residual visualization
• Latent space exploration with UMAP

---

📁 Dataset

• Input: 3D spectral dataset ((samples, 61 wavelengths, 4 components))
• Format: .dat file loaded via NumPy

---

🧠 Model Architecture

Input → [512 → 256] → Latent (128-d) → [256 → 512] → Output (+ Residual)

• Activations: SELU
• Normalization: LayerNorm
• Regularization: Dropout (0.2)
• Loss: α*MSE + β*Cosine + γ*Smoothness

---

🛠️ Installation

pip install torch numpy scikit-learn matplotlib umap-learn

---

▶️ Run Training

python train.py

Where train.py contains:

• Model definition
• Data loading
• Training loop
• Visualizations

---

📊 Visualizations

📉 Training Curves

• MSE loss vs epoch
• Validation tracking with early stopping

🔬 Spectral Reconstruction

• Original vs Reconstructed vs Residuals
• Per component (4 channels)

🌌 Latent Space (UMAP)

• 2D UMAP embedding from the 128-d latent vector

---

📦 Output

• best_deep_res_autoencoder.pth: Saved model with best validation loss

---

📜 License

MIT License © 2025

---

👨‍💻 Author

Dharmik Dudhat
Built using PyTorch + NumPy + Matplotlib
Feel free to ⭐ this repo and contribute!