How apaarraina/cartoon-to-realistic-using-cyclegan works
Project Type and Purpose
• This is a machine-learning/deep-learning research pipeline implementing a cycleGAN to translate cartoon faces (domain A) into realistic human faces (domain B).
• Core purpose: ingest two image domains, train generator & discriminator networks in adversarial fashion, and produce a model that can map cartoons → realistic faces (and vice versa).
File Structure
• Top-level scripts:
• discriminator.py – discriminator network definition
• generator.py – generator network definition
• train.py – standard cycleGAN training script
• train_with_diversity_loss.py – alternative training script with an added diversity loss
• test_realistic_to_cartoon.py – evaluation / inference script for one direction
• train.py/test scripts read/write model checkpoints and images
• preprocessing_scripts/ directory:
• detec_faces.py – face detection on raw images
• filter_frontal_face.py – remove non-frontal faces
• split.py – split cleaned images into trainA, trainB, testA, testB
• Data directories (implied by README):
• cartoonset10k/ – raw cartoon images
• img_align_celeba/ – raw real images
• cartoon2realistic/{trainA, trainB, testA, testB} – post-split image folders
README for Additional Insights
• Mentions pretrained model files (genz.pth, genh.pth, criticz.pth, critich.pth).
• Outlines basic project structure under “cartoon2realistic/”.
Main Components & Their Interactions
a. Data Ingestion Layer
• External Storage (Google Drive links) → Local raw folders: “cartoonset10k/” and “img_align_celeba/”.
b. Preprocessing Layer
• filter_frontal_face.py → keep only good frontal faces
• split.py → shard into trainA, trainB, testA, testB under “cartoon2realistic/”
c. Model Definition Layer
• generator.py → Generator architecture (two networks: G_A2B, G_B2A)
• discriminator.py → Discriminator architecture (D_A, D_B)
d. Training Layer
• train.py / train_with_diversity_loss.py
• Inputs: images from trainA, trainB; network modules; hyperparameters
• Outputs: checkpoint files (genz.pth, genh.pth, criticz.pth, critich.pth)
e. Evaluation/Inference Layer
• test_realistic_to_cartoon.py
• Inputs: test set images from testA or testB, pretrained checkpoints
• Outputs: translated images for qualitative assessment
f. Storage & Artifacts
• Model checkpoint directory
• Generated images directory
• (Optionally) logs and metrics (e.g., loss curves)
Architectural Patterns & Design Principles
• CycleGAN pattern: two generators + two discriminators + cycle consistency losses
• Modular code-separation: distinct scripts for network definitions, training, testing, preprocessing
• Data-pipeline pattern: clear separation of raw data → cleaned data → train/test splits
• Extensibility: alternative training script (diversity loss) demonstrates plug-and-play of new objectives