How cs-chan/artgan works
Project type and purpose
• This is a research‐oriented machine-learning codebase centered on Generative Adversarial Networks (GANs) for artwork and natural image synthesis (“ArtGAN”), plus a companion classification experiment (“ICIP-16”) and the associated WikiArt dataset.
• It bundles (a) data ingestion and preprocessing, (b) custom Python-based GAN model definitions and training scripts, (c) sample generation scripts, (d) evaluation utilities (Inception score), and (e) a set of Caffe‐based fine-tuning experiments for painting classification.
File structure
• Top‐level folders:
• ArtGAN/ – core GAN code
• ICIP-16/ – classification code and paper reference
• “WikiArt Dataset” – raw image files and CSV/txt metadata
• data/ – Caffe mean files and helper scripts for classification
• models/ – Caffe deploy/solver prototxts and train/test shell scripts
• Within ArtGAN/:
• data/ – data_loader.py, multiple ingest_*.py for CIFAR, CUB, Flower, STL, WikiArt; tf_reader for TFRecord ingestion
• nn/ – custom neural-net building blocks (layers, activations, ops)
• models/ – readme (placeholder for future model zoo)
• utils/ – inception_score, utility functions (e.g., saving samples)
• top-level scripts (Artist128GANAE.py, CIFAR64GANAEtrick.py, etc.) – entrypoints for training different conditional GAN variants
• Configuration: requirements.in/requirements.txt for Python dependencies; Caffe prototxts and shell scripts under models/ for classification.
README for additional insights
• Confirms three deliverables: (1) WikiArt Dataset release, (2) “Improved ArtGAN” for conditional synthesis (published in TIP ’19), (3) “Deep Convolutional Network for Fine-art Paintings Classification” (ICIP ’16).
1. Data Sources
– WikiArt images + metadata
– CIFAR-10, CUB-200, Flower-102, STL-10
2. Data Ingestion & Preprocessing
– ingest_*.py scripts → raw images → TFRecords via tf_reader
3. Model Definition
– custom Python nn package (layers, activations, ops)
– GAN architectures (128×128 autoencoder variants)
4. Training Pipeline
– top-level Python scripts (Artist128GANAE.py, etc.)
– hyperparameter configurations
5. Sample Generation
– sample scripts (Artist128GANAEsample.py, etc.)
6. Evaluation
– inception_score module → quality metrics
7. External Tools
– TensorFlow (for TFRecords + training)
– Caffe (for ICIP-16 classification; uses deploy.prototxt, solver.prototxt, shell scripts)
– Data Sources → Data Ingestion → TFRecords → Training Pipeline → Model Checkpoints
– Model Checkpoints → Sample Generation → Generated Images
– Generated Images → Evaluation → Metrics
– For ICIP-16 branch: Data Sources → Caffe mean protos → Caffe Training → Classification Results
c. Architectural patterns & principles
– Modular separation: data, model, training, evaluation, utilities
– Script-based CLI interfaces for each experiment
– Custom neural-net library to decouple model logic from backend (TensorFlow)
– Reuse of external frameworks (TensorFlow for GANs, Caffe for classification)
d. Key technologies & frameworks
– Python 3.x, TensorFlow (implicit via tf_reader), Caffe
– Bash scripts for Caffe training/test
– Custom Python modules (under nn/ and utils/)