Code for paper Prototype-guided Attribute-wise Interpretable Scheme for Clothing Matching.

This project is implemented with

• Python 3.8

• Pytorch 1.6.0

The proposed PAICM is verified on the FashionVC dataset. You can download the origin FashionVC dataset from their provided links (Google Drive Link or Baidu Netdisk Link with the password: ytu4). You also can test the model with your own dataset.

To extract the feature, we provided the item attribute classifier (checkpoint password: a49a) pre-trained on DeepFashion Consumer-to-shop Clothes Retrieval Benchmark to generate the attribute feature. The Consumer-to-shop Clothes Retrieval Benchmark is comprised of 18 attributes and 303 attribute elements. Noted that the provided attribute classifier is a comprehensive one and able to classify all the attributes once a time, which is different with the classifier mentioned in the paper. The ACC of the provided attribute classifier is 91.42%. The feature extraction code is in [fashionVCpredict.py]. The pre-trained attribute classifier can be also used to classify the attributes of other fashion datasets.

To enhance the attribute feature, we also extracted the categoty and color labels from the meta textual and visual data of each item. The extracted categoty and color labels are provided in ./data/. The extraction tools are also provided in ./utils/.

Concat the attribute feature with the categoty and color labels by code [concat_category_color.py], and you can get the final input features.

If you don't want to process the above, you can derictly download the final input features (train_feature.pkl) from the Link with the password: 24nx.

If you have derictly downloaded the final input features, you can skip the following instructions.

• Download the FashionVC dataset.

• Change the path in the head of the fashionVCpredict.py

• Extract the attribute feature with instruction: python fashionVCpredict.py

• Change the path in the head of the concat_category_color.py

• Concat all the feature with instruction: python concat_category_color.py

• python paicm_pytorch.py

We also provided the pre-trained checkpoint for test with the password: cao3.

@inproceedings{HanSYWN19,
  author = {Xianjing Han and Xuemeng Song and Jianhua Yin and Yinglong Wang and Liqiang Nie},
  title = {Prototype-guided Attribute-wise Interpretable Scheme for Clothing Matching},
  booktitle = {Proceedings of the international ACM SIGIR Conference on Research and Development in Informaion Retrieval},
  pages = {785--794},
  year = {2019}
}