Skip to content
No description, website, or topics provided.
Jupyter Notebook Python
Branch: master
Clone or download
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.ipynb_checkpoints sS Dec 8, 2015
.gitignore It worksls Nov 14, 2015
LICENSE Create LICENSE Jan 10, 2016
Readme.md Update Readme.md Jan 10, 2016
acoustic-guitar-player.jpg Finally Works Nov 21, 2015
cnn_util.py test:only tf Dec 8, 2015
ipython_demo.ipynb sss Dec 8, 2015
ixtoword.npy Added files via upload Mar 1, 2016
make_flickr_dataset.py Merge branch 'master' of https://github.com/jazzsaxmafia/show_and_tel… Dec 8, 2015
model.py Update model.py Jan 4, 2016
result.jpg sS Dec 8, 2015

Readme.md

Neural Caption Generator

  • Tensorflow implementation of "Show and Tell" http://arxiv.org/abs/1411.4555
  • Borrowed some code and ideas from Andrej Karpathy's NeuralTalk.
  • You need flickr30k data (images and annotations)

Code

  • make_flickr_dataset.py : Extracting feats of flickr30k images, and save them in './data/feats.npy'
  • model.py : TensorFlow Version

Usage

  • Flickr30k Dataset Download
  • Extract VGG Featues of Flicker30k images (make_flickr_dataset.py)
  • Train: run train() in model.py
  • Test: run test() or test_tf() in model.py
  • parameters: VGG FC7 feature of test image, trained model path
  • Once you download Tensorflow VGG Net (one of the links below), you don't need Caffe when testing.

Downloading data/trained model

  • Extraced FC7 data: download
  • This is used in train() function in model.py. You can skip feature extraction part by using this.
  • Pretrained model download
  • This is used in test() and test_tf() in model.py. If you do not have time for training, or if you just want to check out captioning, download and test the model.
  • Tensorflow VGG net download
  • This file is used in test_tf() in model.py
  • Along with the files above, you might want to download flickr30k annotation data from link

alt tag

License

  • BSD license
You can’t perform that action at this time.