How morancium/image_colorization works
Project Type and Purpose
• This is a deep‐learning pipeline for automatic image colorization.
• It uses PyTorch and a cGAN (generator + PatchGAN discriminator) with two generator variants (custom U-Net and ResNet18-backbone U-Net).
• Key workflows: training from scratch, training with pretrained backbone, and inference on new images.
File Structure
• config.py – hyperparameters (LR, batch size, epochs, image size, L1 loss weight).
• dataset.py – dataset class & DataLoader (downloads/loads COCO or custom images, converts RGB → L*a*b*).
• generator_model.py – custom U-Net definition.
• pretrained_generator.py – ResNet18-based U-Net definition.
• PatchDiscriminator_model.py – PatchGAN discriminator implementation.
• train.py – training loop orchestrating generator vs. discriminator updates.
• pretrained_model.py – alternate training loop using the pretrained generator.
• test.py – inference script (loads grayscale images, runs generator, recombines L + predicted ab).
• files/ – likely for storing downloaded dataset or interim files.
Key Components to Showcase
a. Data Ingestion Layer
• COCO Dataset or Custom Images → dataset.py → transforms (resize, tensor, L*a*b* conversion).
b. Preprocessing → splitting L channel (input) vs. a/b channels (ground truth).
c. Model Definitions
• Generator A: Custom U-Net (generator_model.py)
• Generator B: ResNet18 U-Net (pretrained_generator.py)
• Discriminator: PatchGAN (PatchDiscriminator_model.py)
d. Training Orchestrator
• train.py or pretrained_model.py uses config.py parameters → loops:
1) Discriminator update (real vs. fake patches, BCE loss)
2) Generator update (adversarial BCE + L1 pixel loss)
• Visualization & checkpointing (utils.py functions save example outputs and model weights into evaluations/ or output/).
e. Inference Pipeline
• test.py takes new grayscale images → loads checkpoint → generator → recombines channels → outputs colorized images.
f. Supporting Utilities
• config.py (tunable hyperparameters)
• utils.py (loss functions, image I/O, Lab conversions)
Relationships & Data Flows
• Output of Generator → L1 Loss & Adversarial Loss comparators → back to training orchestrator
• Save checkpoints & sample images → Evaluation directory
• Inference branch off the Generator → final colorized images
Architectural Patterns & Principles
• Modular separation: data layer, model definitions, training logic, inference logic, configuration, utilities
• cGAN pattern: two cooperating adversarial networks
• U-Net skip connections for fine‐grained pixel mapping
Final Checklist
• Clear labels on all components and files
• External dependencies:
– PyTorch & CUDA (GPU)
– COCO dataset (or custom dataset input)
• Hyperparameter configuration from config.py
• Two training paths (from-scratch vs. pretrained) converging on evaluation outputs