How ilyes2000/unpaired-image-to-image-translation-using-cycle-consistent-adversarial-networks works
Project Type & Purpose
• This is a research-style deep learning pipeline implementing CycleGAN for unpaired image-to-image translation.
• It is organized as a CLI-driven training/evaluation tool, with modular data loading, model definitions, training loops, checkpointing, logging, and testing.
File Structure
• Top-level scripts
– train.py / train_from_checkpoint.py: entry points for training
– test.py / evaluate_fcn.py: evaluation & inference
• Configuration
– options/base_options.py, train_options.py, test_options.py: define and parse CLI args
• Data handling
– data/dataset.py, data/unaligned_dataset.py: load & preprocess two unpaired image domains
• Models
– models/generators.py: generator architectures
– models/discriminators.py: discriminator architectures
– models/networks.py: shared network blocks
– models/cycle_gan_model.py: orchestrates generator + discriminator + cycle losses
• Utilities
– utils.py, metrics.py: helper functions, loss calculations, logging, plotting
• Persistence
– model_checkpoints/: saved .pth files per epoch
– checkpoints/: raw logs for custom “horse2zebra” runs
– results/, samples/: CSV/visual output of experiments
Main Components & Their Interactions
a. CLI / Options Parser
– Parses user flags (epochs, lr, lambda_cycle, device).
– Feeds config into Training Orchestrator and Data Loader.
b. Data Loader
– Reads two image domains from disk via UnalignedDataset
– Applies transforms, batches data.
c. CycleGAN Model Orchestrator
– Instantiates two Generators (G_AB, G_BA) and two Discriminators (D_A, D_B).
– Coordinates forward passes:
• Real A → G_AB → Fake B → D_B → adversarial loss
• Real B → G_BA → Fake A → D_A → adversarial loss
• Cycle: Fake B → G_BA → Rec A; Fake A → G_AB → Rec B
– Computes cycle-consistency, identity, adversarial losses.
d. Optimizers & Schedulers
– Two optimizer groups: one for generators, one for discriminators.
– LR schedulers applied per epoch.
e. Training Loop
– For each epoch / batch:
1. Load real A & B
2. Forward passes through G, D
3. Compute and backprop losses
4. Step optimizers, update schedulers
5. Log scalar losses and images via utils.
f. Checkpointing & Logging
– Saves model weights (.pth) in model_checkpoints/ after specified epochs.
– Exports loss logs (CSV) to checkpoints/ and results/.
g. Evaluation & Visualization
– evaluate_fcn.py / test.py load saved checkpoints, run inference on test sets.
– Outputs samples/ images and fcn_scores.csv.
Architectural Patterns & Principles
• Modular separation: CLI, Data, Model, Training, Evaluation modules.
• Dependency injection via options classes.
• Clear pipeline: Data → Model → Loss → Optimizer → Checkpoint/Log.
• Extensibility: swap datasets, architectures or hyperparams via options files.
1. Sketch top-level modules in a horizontal flow: CLI → Data Loader → Training Engine.
2. Drill into Training Engine: show Generators ↔ Discriminators ↔ Cycles.
3. Side-paths: Checkpoint writer, Logger.
4. A separate branch for “Evaluation” loading saved models.