This is an implementation for the paper: https://arxiv.org/abs/1803.08024.
master branch: t2i model as described in the paper. i2t branch: i2t model as described in the paper.
To train the model
python main.py --mode 0
To load and test the best model
python main.py --mode 1
Note: Look up main.py for passing additional arguments while training.
Results on fast branch which is a faster version of master and is a t2i model:
| r@1 (t2i) | r@5 (t2i) | r@10 (t2i) | r@1 (i2t) | r@5 (i2t) | r@10 (i2t) |
|---|---|---|---|---|---|
| 0.48314 | 0.7674 | 0.824 | 0.624 | 0.878 | 0.929 |