Deep_Hierarchical_Classification

A non-official PyTorch implementation of Deep Hierarchical Classification for e-commerce category prediction, using a ResNet50 with CBAM backbone and tested on Cifar-100.

The README is careful to call this a non-official implementation of the DHC paper. That means it was written by the repository author rather than released by the original researchers, and the code is a working reference for the approach. The project description frames the target as category prediction in e-commerce systems. Because it is a reimplementation, the code tracks the paper's structure rather than adding new machinery.

The classifier uses ResNet50 with CBAM as its backbone network. Performance was tested on Cifar-100, and the README describes the dataset structure in detail so the hierarchy is clear before any custom work starts. The backbone pairing is simply stated, without an explanation of why the two are combined.

Cifar-100 is arranged in two levels. There are 20 superclasses, and each superclass holds 5 subclasses, giving 100 subclasses in total. Each subclass has 600 images, split as 500 for training and 100 for testing. That two-level shape is exactly what the hierarchical classifier is built to exploit.

To use the code on a custom dataset, the README asks for train and test CSV files plus a metafile, with the dataset folder serving as the reference. The dictionary variable in the level dict file must be updated to match the dataset hierarchy, and the load dataset file has to be modified as well. Datasets with more than two levels require adding classifier layers in the ResNet50 model file. The steps are ordered, so a two-level dataset needs only the first two changes. The README gives no accuracy numbers or training recipes beyond this.

• Source 1 · Project repository
• Source 2 · README