w2vv_py35_keras_tensorflow

transfer the py2-version repository w2vv (official code https://github.com/danieljf24/w2vv) to py3-version code and test on Windows 10

Changes

1. adjust the print function
2. solve the encoding problem of reading word2vec id.txt and building dict in text2vec.py
3. import cPickle as pkl to import _pickle as pkl
4. open(".pkl", "r") to open(".pkl", "rb")
5. all map(x,xxx) function to list(map(x,xxx))
6. rank = np.where(inds/n_caption == i)[0][0] in i2t function in evaluation.py to rank = np.where(inds//n_caption == i)[0][0], since all 5 captions of the same image are taken into evluation
7. etc ...

Enviroment

1. Windows 10
2. python 3.5
3. TensorFlow 1.8.0 Cuda 9.0
4. keras 2.1.6 with TensorFlow backend
5. pydot 1.2.3 + Graphviz
6. tensorboard_logger for tensorboard visualization

Demo train and test on flickr8K with default setting

1. My test : Image to text: recall@1 34.2, recall@5 64.9, recall@10 77.3, 3.0, 13.6
2. Paper Report: Image to text: recall@1 36.3, recall@5 66.4, recall@10 78.2