How amlab-amsterdam/diva works
Project type and purpose
• This is a research‐oriented PyTorch codebase implementing and comparing several domain‐adaptation and generative models (CNN, DANN, VAE, DIVA, conditional generators) on two datasets: malaria cell images and rotated MNIST. The main goal is to train and evaluate domain‐invariant variational autoencoders (DIVA) and variants, plus baselines.
Main components
a. Data module
• Datasets: “paper_experiments/malaria/dataset” and “paper_experiments/rotated_mnist/dataset”
• Data loaders (mnist_loader.py, data_loader_topk.py, etc.)
• Sup/unsup splits and additional unlabeled domains
b. Model module
• CNN variants under malaria/cnn
• Domain‐adaptation models under malaria/dann and rotated_mnist/additional_unlabeled_domain
• Generative models: VAE (vae/), DIVA (diva/) with and without z_d
• Utility code for model blocks (resnet blocks, pixel_cnn_utils)
c. Experiment scripts
• Individual .py scripts that set up configurations, initialize data loaders, instantiate models, define losses, run training loops, and save checkpoints or outputs
• Organized by dataset & method (e.g., experiment_diva_top10_resnet_new.py, experiment_vae_top10_resnet_new.py)
d. Output & evaluation
• Generated images (conditional_generation/generate_images.py)
• Trained model checkpoints (trained_models directories)
• Test‐set evaluation scripts (calculate_test.py)
e. External frameworks
• PyTorch 1.x, scikit‐image, scikit‐learn, seaborn
• Python 3.6
Relationships and interactions
• Experiment script imports Data module to fetch batches
• Batches fed into Model module
• Model’s forward pass → compute losses (reconstruction, domain classification, KL divergence)
• Backpropagation via PyTorch optimizer
• At checkpoints: save model weights, optionally call conditional generation utility
• Evaluation scripts load saved checkpoints, run inference, write metrics
Architectural patterns and principles
• Modular separation of data loaders, model definitions, and experiment orchestration
• Reusable building blocks (resnet blocks, VAE encoders/decoders, domain‐classifier heads)
• Clear directory‐by‐method organization encourages plug‐and‐play comparisons
• Scripts follow an “initialize → train → save → evaluate” pipeline