How omid-ghozatlou/query-by-example-using-svdd works
Project type and purpose
• This is a research‐oriented deep learning toolkit for “query by example” image retrieval in remote sensing archives, built around an enhanced Deep SVDD (Support Vector Data Description) framework.
• Key capabilities: multiple neural architectures (Gabor CA ResNet, Split Transform Network, various LeNets), a TCAL (Triple Criteria Active Learning) extension, dataset loaders/preprocessors, training pipelines, and evaluation/visualization tools (t-SNE, histograms).
File structure
• Top-level scripts / entry points:
• main.py, script.py – generic runners
• deepSVDD.py – core SVDD class
• TCAL.py – active‐learning extension
• Gabor_CA_ResNet.py, Split_Transform_Network.py – network definitions
• IR_evaluation.py, tsne.py, histogram.py – evaluation/analysis
• base/ – abstract base classes for datasets, networks, trainers
• datasets/ – each file handles a dataset (CIFAR-10, EuroSAT, MNIST, etc.) + preprocessing
• networks/ – concrete network architectures ready to plug into DeepSVDD or TCAL
• optim/ – trainer implementations (ae_trainer, deepSVDD_trainer)
• utils/ – config loading, result collection, visualization helpers
• notebooks/ – example Jupyter notebooks demonstrating usage
Extract insights from the README
• The workflow:
a. Load images via dataset modules
b. Preprocess & augment
c. Instantiate a network (e.g., SplitTransformNetwork)
d. Wrap it in DeepSVDD (or TCAL for active learning)
e. Train via optim/*.py trainers
f. Retrieve/query images and evaluate with IR_evaluation, t-SNE, histograms
• Tech stack: Python 3.7+, PyTorch, NumPy, scikit-learn, Matplotlib
Components & relationships
a. Main components (blocks):
1. Data Sources – CIFAR-10, EuroSAT, MNIST, custom datasets
2. Preprocessing Pipeline – normalization, augmentations, transforms
3. Network Architectures –
• Gabor Convolutional Attention ResNet
• Split Transform Network
• LeNet variants (cifar10_LeNet, scatter_LeNet*)
4. Model Wrappers – DeepSVDD, TCAL
5. Trainers – AETrainer, DeepSVDDTrainer
6. Retrieval/Evaluation – IR_evaluation, tsne, histogram
7. Visualization – plot_images_grid, t-SNE scatter plots, histograms
8. Configuration & Utilities – config parsing, results logging
9. Entry Points – main.py, script.py, notebooks
b. Data flow & interactions:
• Data Sources → Preprocessing → Network Architectures
• Networks + hyperparameters → wrapped by Model Wrappers
• Model Wrappers + Trainers → Training Loop → Checkpoints / Metrics
• Trained Model + Query Image → Retrieval Module → Ranked Results
• Results → Evaluation Scripts (IR_evaluation, tsne, histogram)
• Evaluation Outputs → Visualization Module → Plots
c. Architectural patterns / principles:
– Modular design via base classes (open/closed principle): plug‐in new datasets, nets, trainers
– Separation of concerns: data loading, model definition, training logic, evaluation, visualization
– Configurable pipelines via utils/config.py
– Reproducibility: clear script/notebook entry points
d. Key technologies & frameworks:
– PyTorch (nn.Modules in networks/, training in optim/)
– NumPy / scikit-learn (some preprocessing, metrics, t-SNE)
– Matplotlib (visualization)
– Python scripts + Jupyter notebooks for interactive demo
Final advice
• Then expand each stage into subcomponents, reflecting the directory structure.
• Review with the team to ensure that extension points and dependency flows are correctly captured.