Outfit-Transformer
🤗 Introduction
Implementation of paper - Outfit Transformer: Outfit Representations for Fashion Recommendation
⚠️ The original paper outlines the specifics of the target item for Compatitible Item Retrieval (CIR) and Fill-in-the-Blank (FITB). Nonetheless, for the sake of impartial evaluation alongside other models, this information was intentionally excluded. (Should a dataset emerge that necessitates the prediction of a matching item when presented with a description unrelated to the target item itself, the model will be retrained accordingly.)
🎯 Performance
The figures below are derived using the Polyvore test dataset.
|Model|CP(AUC)|FITB(Accuracy)|CIR(Recall@10)|
|:-|-:|-:|-:|
|Type-Aware|0.86|57.83|3.50|
|SCE-Net|0.91|59.07|5.10|
|CSA-Net|0.91|63.73|8.27|
|OutfitTransformer(Paper)|0.93|67.10|9.58|
|**Implemented
(w/o target desc.)**|**0.91**|**64.10**|Not Trained|
(w/o target desc.)**|**0.91**|**64.10**|Not Trained|
⚙ Install Dependencies
This code is tested with python 3.9.16, torch 1.12.1
python -m pip install -r requirements.txt🧱 Train
Data Preparation
Download the polyvore dataset from here
Pretraining on CP(Compatibiliby Prediction) task
python train.py --task cp --train_batch 64 --valid_batch 96 --n_epochs 5 --learning_rate 1e-5 --scheduler_step_size 1000 --work_dir $WORK_DIR --data_dir $DATA_DIR --wandb_api_key $WANDB_API_KEYFinetuning on CIR(Complementary Item Retrival) task
python train.py --task cir --train_batch 64 --valid_batch 96 --n_epochs 5 --learning_rate 1e-5 --scheduler_step_size 1000 --work_dir $WORK_DIR --data_dir $DATA_DIR --wandb_api_key $WANDB_API_KEY --checkpoint $CHECKPOINT🔍 Test
python test.py --task $TASK --polyvore_split nondisjoint --test_batch 96 --data_dir $DATA_DIR --checkpoint $CHECKPOINT🧶 Checkpoints
Download the checkpoint from here
🔔 Note
- A paper review of implementation can be found at [here](). (Only Available in Korean)
- This is NON-OFFICIAL implementation. (The official repo has not been released.)