Customised CIFAR-10 Dataset for minimal running. Accepts dual_input parameter to enable dual image output. Note that images are identically duplicated only.
- VGG16 series
- VGG11 series
- ReducedVGG11 series
- STCNN (Spatial Transformer CNN) series
Each series has Dual or not version.
VGG series mainly exists for basic pipeline purpose and experimentation.
STCNN implements from the paper Spatial Transformer Networks, which adds spatial invariance (any affine transformation such as rotation, scaling, shearing) property to the network. Spatial Transformer blocks are easily pluggable to any neural network.
A basic pipeline that is almost self-explanatory. It has "setting-train-testing-save" workflow, but normally only "setting-testing" is used. Settings are minimal, no hyper-parameter optimisation. This serves as template for any expansion to a real experiment.
The arguments are inside run() and encouraged to modify variables in the space provided at the beginning of the function, instead of wrapping them as arguments of the function.
- No suitable dataset
- Not trained models
- Not exhaustive review