README.md

FashionIQChallenge2020

2nd Place Team's Solution for Fashion IQ Challenge 2020

• CodaLab: https://competitions.codalab.org/competitions/22946#results
• Workshop page: https://sites.google.com/view/cvcreative2020/fashion-iq

(1) Environment Setup

• Prepare Data:
  Download Dataset from here

Image download script has been added. (12/07/2020)

run_download_image.sh

The data structure should look like:

cd ./ours/train
dataset
   ├──image_splits
   │       ├──split.dress.train.json
   │       ├──...
   │       ├──split.toptee.test.json
   │
   ├──captions
   │       ├──cap.dress.train.json
   │       ├──...
   │       ├──cap.toptee.test.json
   │
   ├──images
           ├──B000OHM9FI.jpg
           ├──...
           ├──B000M89C40.jpg

• Install required packages using virtualenv

python3 -m virtualenv --python=python3 py3
. py3/bin/activate
pip install -r requirements.txt

• Download required files: word2vec, sentence_embedding, image_emedding
  Download from here: google drive

cd ./ours/train
tar -xvf assets.tar

(2) Train a Single Model (TIRG, Best Score: 37.18)

Run the script below.

overall score curve

loss curve

cd ours/train
python3 main.py \
    --warmup \
    --gpu_id '0' \
    --method 'tirg' \
    --text_method 'lstm-gru' \
    --expr_name 'devel' \
    --data_root './dataset' \
    --backbone 'resnet152' \
    --fdims 2048 \
    --epochs 100 \
    --batch_size 32 \
    --image_size 224 \
    --normalize_scale 5.0 \
    --lr 0.00011148 \
    --lrp 0.48 \
    --lr_decay_steps "10,20,30,40,50,60,70" \
    --lr_decay_factor 0.4747

(3) Ensemble Scores using Bayesian Optimization

(i) get the score output on test/val:

cd ours/tools
python3 get_score.py --data_root './dataset' --expr_name 'devel'

(ii) get emsemble score using Bayesian Optimization provided by hyperopt:
Check the name of output score saved in ours/tools/output_score.
if the candidate models are: ['20200605122245_devel', '20200605152325_devel']

python3 optimize_score.py --data_root './dataset' --repos '20200605122245_devel,20200605152325_devel'

Authors

• Minchul Shin, Search Solutions
• Yoonjae Cho, NAVER/LINE Vision
• Seongwuk Hong, NAVER/LINE Vision