How tinhvu11235/unet_colorization works
Project Type & Purpose
• Primary goal: train the UNet (with checkpointing via WandB) and provide a lightweight demo inference script.
Main Components
a. Configuration Layer
• config.py – defines hyperparameters and dataset paths for scratch vs. resume modes.
b. Data Ingestion
• data_loader.py – reads images from TRAIN/VAL folders, applies transformations, outputs PyTorch DataLoader objects.
c. Model Definition
• model.py – implements the UNet architecture for Lab-space colorization.
d. Training Orchestration
• train.py – runs the training loop (forward/backward), scheduler, optimizer, and periodically saves checkpoints.
• main.py – entry point: parses CLI args (--mode scratch or --mode continue), loads config, invokes WandB artifact loading for resumes, then calls train.py.
e. Checkpointing & Experiment Tracking
• Weights & Biases (WandB) – used to log metrics, store checkpoints as artifacts, and retrieve latest checkpoints when resuming.
f. Inference / Demo
• demo_release.py – lightweight script that pulls the latest model artifact, loads the trained UNet, accepts a grayscale image, outputs a colorized image.
External Services & Storage
• Dataset Storage (local or network file system)
• WandB Cloud for run tracking and artifact storage
• Local filesystem for temporary downloads (artifact_dir) and output images
Relationships & Data Flow
[config.py] –> feeds hyperparameters to → [main.py]
[main.py] –> loads WandB checkpoint (if continue mode) → downloads artifact → passes model & optimizer states to → [train.py]
[config.py] –> provides paths to → [data_loader.py] –> yields batches to → [train.py]
[model.py] –> defines network used by → [train.py]
[train.py] –> pushes checkpoints & metrics to → [WandB Cloud]
Inference Flow (Bottom):
[demo_release.py] –> uses WandB API to pull latest model artifact → loads UNet from [model.py] → processes input grayscale image → writes colorized output to disk
Architectural Patterns & Design Principles
• Modular separation: config, data, model, training, inference
• Single responsibility per script
• Cloud-native experiment tracking & artifact management via WandB
• CLI-driven entry points for reproducibility
Technologies & Frameworks
• Python 3.x, PyTorch, torchvision
• numpy, scikit-image, matplotlib, tqdm
• Weights & Biases (wandb)
• gdown (for optional dataset download)