TIPCB

Prerequisites

• Pytorch 1.1
• cuda 9.0
• python 3.6
• GPU Memory>=12G

Datasets

We evaluate our method on CUHK-PEDES. Please visit Here.

Usage

• You need to generate tokens to "/data/BERT_encode/" by running "BERT_token_64.py" or downloading from Here.
• If you want to train the network, you can run our code with the following commands:

python train.py --max-length 64 --batch-size 64 --num-epoches 80 --adam-lr 0.003 --gpus 0

• You can download our trained model from Here and the trained log from Here.

Evaluate

| Top-1 | Top-5 | Top-10 | | :------: | :------: | :------: | | 63.63 | 82.81 | 89.01 |