LCPN-hiernet is a hierarchical image classification model for fashion commerce items based on EfficientNet-b4 and LCPN (Local Classifier per Parent Node) technique. This particular implementation is designed for level-2 hierarchy datasets, but I'm working on making it available for deeper datasets in the future!

Local classifier per parent node: training one multi-class classifier for each parent node, to distinguish between its child nodes.

In our example, that would mean one classifier on the first level (to determine “cat”, “dog” or “unicorn”), then three more classifiers to determine the specific breed.

In order to prevent inconsistencies with the predictions of different levels, you can design the system such that an example classified by the first classifier as “cat” would only be seen by the “Cats” classifier at the second level. This way, you can avoid the creation of new, exotic hybrids such as a Poodle Unicorn (though that does sound marvelous)."

-- The Hitchhiker’s Guide to Hierarchical Classification

$ python3 -m venv venv
$ source ./venv/bin/activate
$ pip3 install -r requirements.txt

First, prepare a level-2 hierarchy dataset organized in the following structure.

./dataset
├── parent_class_1
│   ├── child_class_1
│   │    ├── img_0001.jpg
│   │    ├── img_0002.jpg
│   ├── child_class_2
│   │    ├── img_0001.jpg
│   │    ├── img_0002.jpg
├── parent_class
│   ├── child_class_3
│   │    ├── img_0001.jpg
│   │    ├── img_0002.jpg
│   ├── child_class_4
│   │    ├── img_0001.jpg
│   │    ├── img_0002.jpg
│   │    ├── etc. . .

Refer to dataset.example for the directory structure used for our specific application of classifying CHxNEL e-commerce products.

Then, run the following script:

$ python3 prepare_dataset.py -p ./dataset_example

This script will prepare all datasets to train each local classifier under split_datasets folder. For our dataset_example, the script created 7 different datasets split into training and validation set; 1 parent classification model and 6 child classification models.

$ cd fashion-net/src
$ chmod +x ./train_lcpn.sh
$ ./train_lcpn.sh

This script will train all models consecutively from datasets in split_datasets folder. All of following output files will be saved in fashion-net/trained_models folder.

trained_models
├── accessories
│   ├── bin
│   │   ├── encoder.pickle
│   │   ├── inv_normalize.pickle
│   │   └── test_transforms.pickle
│   ├── class_plot.png
│   ├── confusion_matrix.png
│   ├── error_plot.png
│   ├── model.pth
│   ├── performance.txt
│   └── wrong_plot.png
├── bags
├── cosmetics
├── parent
├── perfume
├── wallets
└── watches

$ cd fashion-net/src/
$ python3 predict.py -p RELATIVE_PATH_TO_IMAGE

Pass in the path to the image you wish to predict on, and this script will perform a two-stage prediction: Parent class & Child class.

1. Parent Class Prediction -> Classified as "bags"
2. Load encoder and model for "bags" child_class
3. Child Class Prediction -> Classified as "bags_boy"

$ python3 predict.py -p ../sample_images/bags_b.jpg
=======Phase 1: Parent Class Prediction========
Loading pkl files...
ENCODER:  {0: 'accessories', 1: 'bags', 2: 'cosmetics', 3: 'perfume', 4: 'wallets', 5: 'watches'}
Loading model.pth...
Forward Passing...
Predicting Label...
Parent class prediction:  bags
=======Phase 2: Child Class Prediction========
Loading pkl files...
ENCODER:  {0: 'bags_a', 1: 'bags_b', 2: 'bags_c', 3: 'bags_d', 4: 'bags_e', 5: 'bags_f', 6: 'bags_g'}
Loading model.pth...
Forward Passing...
Predicting Label...
Final Prediction:  bags_b

Avg. Acc: 85.19%

MIT