DL Learning Lab

Deep learning study notes and project code — learning bottom-up through hands-on experiments.

Structure

DL-Learning-Lab/
├── _DL_Notes/          # Obsidian notes (theory + experiment analysis)
│   └── _02_Architectures/
│       ├── DDPM-扩散模型入门.md
│       ├── SegFormer-语义分割入门.md
│       ├── Stable-Diffusion-LoRA入门.md
│       └── ...
├── projects/           # Source code for each project
│   ├── ddpm/           # DDPM conditional generation (MNIST + CFG)
│   ├── segformer/      # SegFormer from-scratch (MiT-B0, VOC2012)
│   ├── segformer-voc/  # SegFormer fine-tuning (HuggingFace, VOC2012)
│   ├── srcnn/          # SRCNN super-resolution (VOC2012)
│   ├── zero-dce/       # Zero-DCE low-light enhancement
│   ├── deraining/      # SimpleMPRNet image deraining
│   ├── aodnet/         # AOD-Net image dehazing
│   ├── resnet/         # ResNet classification
│   ├── deblurring/     # Image deblurring
│   ├── denoising/      # Image denoising
│   └── yolo-football/  # YOLOv11s football detection
└── colab/              # Colab notebooks (SD LoRA, etc.)

Projects

Project	Task	Dataset	Key Result
ddpm	Conditional image generation	MNIST	CFG digit generation (0-9)
segformer-voc	Semantic segmentation (fine-tune)	VOC2012	mIoU = 0.6658
segformer	Semantic segmentation (from scratch)	VOC2012	mIoU = 0.1367 (epoch 40)
srcnn	Super-resolution	VOC2012	PSNR = 26.32 dB
zero-dce	Low-light enhancement	VOC2012	PSNR delta +0.80 dB
deraining	Image deraining	VOC2012	PSNR = 30.60 dB
aodnet	Image dehazing	VOC2012	—
yolo-football	Object detection	Football dataset	mAP50 = 0.816

Notes

Study notes are written in Obsidian and cover architecture theory, experiment logs, and latest research.
Each note links to the corresponding code in projects/.

Model weights (.pth, .pt) and datasets are excluded from this repo (too large).