- Anaconda 3
- Extra Python modules
- Keras >= 2.0.4
- TensorFlow >= 1.1.0
- tqdm
- pycoco (fixed for python3)
The script train_td.py
uses MSCOCO dataset placed at ./COCO
as follows.
./COCO
├── annotations
│ ├── captions_train2014.json
│ ├── captions_val2014.json
│ ├── image_info_test-dev2015.json
│ ├── image_info_test2014.json
│ ├── image_info_test2015.json
│ ├── instances_train2014.json
│ ├── instances_val2014.json
│ ├── person_keypoints_train2014.json
│ └── person_keypoints_val2014.json
└── images
├── test2014
├── test2015
├── train2014
└── val2014
- git-lfs
# for training
python train_td.py
For caption generation demo, see test_predict.ipynb
.