This area contains image generation, classification, object detection, and
robustness experiments implemented with TensorFlow model components and local
training scripts. Families are organised by architecture rather than by a
shared framework; read each folder's config.py before running it because
dataset roots and runtime controls are local.
| Family | Scope | Status | Documentation |
|---|---|---|---|
| Diffusion | Ten numbered denoising/generative-model folders | Mixed: source exists; several dataset/train entrypoints are explicit scaffolds | Open |
| GANs | Eleven adversarial image-generation variants | Source-backed, configuration-driven training folders | Open |
| Image classification | Small/large CNNs, compression/distillation, ViT, and robustness scripts | Architecture code and local dataset/training paths | Open |
| Object detection | R-CNN variants, Mask R-CNN, YOLOv1–v3, and SSD | Detector/train/test code with COCO, VOC, and ImageNet-related adapters | Open |
The source uses tf.keras building blocks extensively, with custom
tf.keras.layers.Layer and tf.keras.Model classes for several generative and
detection systems. Custom optimisation paths use tf.GradientTape; datasets
use tf.data where streaming and augmentation are needed. Selected folders
configure MirroredStrategy when multiple GPUs are visible, while ordinary
Keras execution remains usable on a single GPU or CPU.
Training outputs—including logs, checkpoints, samples, TensorBoard events, and cached detector targets—are local artifacts. They are not evidence of a reproducible benchmark in the tracked source and should not be committed.