Tensorflow implementation of soft-attention mechanism for video caption generation.
An example of soft-attention mechanism. The attention weight alpha indicates the temporal attention in one video based on each word.[Yao et al. 2015 Describing Videos by Exploiting Temporal Structure] The original code implemented in Torch can be found here.
- Python 2.7
- Tensorflow >= 0.7.1
- NumPy
- pandas
- keras
- java 1.8.0
The MSVD [2] dataset can be download from here.
We pack the data into the format of HDF5, where each file is a mini-batch for training and has the following keys:
[u'data', u'fname', u'label', u'title']
batch['data']
stores the visual features. shape (n_step_lstm, batch_size, hidden_dim)
batch['fname']
stores the filenames(no extension) of videos. shape (batch_size)
batch['title']
stores the description. If there are multiple sentences correspond to one video, the other metadata such as visual features, filenames and labels have to duplicate for one-to-one mapping. shape (batch_size)
batch['label']
indicates where the video ends. For instance, [-1., -1., -1., -1., 0., -1., -1.]
means that the video ends at index 4.
shape (n_step_lstm, batch_size)
We generate the HDF5 data by following the steps below. The codes are a little messy. If you have any questions, feel free to ask.
Once you change the video_path
and output_path
, you can generate labels by running the script:
python hdf5_generator/generate_nolabel.py
I set the length of each clip to 10 frames and the maximum length of frames to 450. You can change the parameters in function get_frame_list(frame_num)
.
label_path
: The path for the labels generated earlier.
feature_path
: The path that stores features such as VGG and C3D.
You can change the directory name whatever you want.
h5py_path
: The path that you store the concatenation of different features, the code will automatically put the features in the subdirectory cont
python hdf5_generator/input_generator.py
Note that in function get_feats_depend_on_label()
, you can choose whether to take the mean feature or random sample feature of frames in one clip. The random sample script is commented out since the performance is worse.
I set the maxmimum number of words in a caption to 35. feature folder
is where our final output features store.
python hdf5_generator/trans_video_youtube.py
(The codes here are written by Kuo-Hao)
video_data_path_train = '$ROOTPATH/SA-tensorflow/examples/train_vn.txt'
You can change the path variable to the absolute path of your data. Then simply run python getlist.py
to generate the list.
P.S. The filenames of HDF5 data start with train
, val
, test
.
$ python Att.py --task train
Test the model after a certain number of training epochs.
$ python Att.py --task test --net models/model-20
We modified the code from this repository jazzsaxmafia/video_to_sequence to the temporal-attention model.
[1] L. Yao, A. Torabi, K. Cho, N. Ballas, C. Pal, H. Larochelle, and A. Courville. Describing videos by exploiting temporal structure. arXiv:1502.08029v4, 2015.
[2] chen:acl11, title = "Collecting Highly Parallel Data for Paraphrase Evaluation", author = "David L. Chen and William B. Dolan", booktitle = "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics (ACL-2011)", address = "Portland, OR", month = "June", year = 2011