t-SNE visualisation of CNN features can help understand how the CNN understands the relationship between different classes which it is trained on. This repository contains minimal code to visualise t-SNE embeddings of the features obtained after average pooling in the ResNet-18 architecture. For visualisation the validation set of Imagenette was used.
- The jupyter notebook generates .tsv files containing the features (vis.tsv) and the corresponding labels (metadata.csv)
- Upload both files to : https://projector.tensorflow.org/
- Et voila!
