Others#
Domain-agnostic method demonstrations: image and video recognition, domain adaptation, few-shot learning and data loading. Start here if you are new to PyKale. 10 tutorials.
Draft
Transfer an action recogniser between egocentric video domains. Same adversarial adaptation pipeline as the digits tutorial, with a 3D-CNN feature extractor and video data loaders.
Draft
Classify spoken-and-written digits from paired audio and image inputs, comparing late fusion, low-rank tensor fusion and bimodal interaction fusion on AVMNIST.
Draft
Prototypical Networks under N-way K-shot settings. Learn an embedding where class prototypes are the mean of their support samples, then classify query samples by distance to those prototypes.
Draft
Train an 8-layer CNN on CIFAR-10 as a baseline, then stack a Transformer encoder on the same backbone and compare validation accuracy across four Transformer sizes.
Draft
Deep isometric learning on CIFAR-10/100. ISONet trains very deep networks without residual connections or batch normalisation by keeping the layer-wise transform close to an isometry.
Written
Adapt a digit classifier from MNIST to USPS without target labels, comparing adversarial and discrepancy-based methods. The reference tutorial for domain adaptation in PyKale.
Draft
Adapt from three labelled source domains to one unlabelled target using moment matching (M3SDA) and multiple feature-space alignment (MFSAN), on Office-Caltech and digits.
Draft
Autism detection from resting-state fMRI pooled across four ABIDE acquisition sites, using CoIRLS to regularise away site-driven covariate shift.
Draft
The smallest possible domain adaptation example. Two blob-generated domains, two classes, and a CoIRLS classifier – runs in seconds and makes the covariate-shift problem visible in one plot.
Draft
Load, sample and augment video datasets with VideoFrameDataset. Sparse temporal sampling keeps memory flat regardless of clip length, and the loader plugs straight into torch DataLoader.