/image_captioning

Image Captioning with Transfer Learning, Attention, Pretrained Embeddings, and Teacher Forching Techniques

Primary LanguagePython

Image Captioning with Transfer Learning, Attention, Pretrained Embeddings, and Teacher Forcing Techniques

Installing

Python Dependencies

Using Docker

You can create docker image using Dockerfile

    $ cd Docker
    $ docker build -t im_caption:v1 .

then you can create docker container

    $ docker run -it --rm -v `pwd`:/workspace im_caption:v1

then you can run your python commands from there

Using Conda

  • First download pytorch and torchvision from https://pytorch.org using conda
  • Then, you can download required libraries using pip inside Docker/Dockerfile

Data Set

Selected dataset was flicker 30k in format of .h5. (Right now I am trying to provide a download link)

  • This file includes urls for the images. To download images from those urls run data_extractor.py

  • dataset includes image ids, integer captions, word captions. You need to create .csv files

    • These files should be created captions_words.csv, captions.csv, imid.csv and word2int.csv
    • (Right now I will try to provide the code for creation of .csvs)

Methodology

Transfer Learning

For CNN architecture on image feature extraction pretrained VGG16 is used,

self.vgg = models.vgg16(pretrained=self.pre_train)

Until 15th epoch gradients of VGG16 is freezed. After Epoch 15 last layers activated

Attention

Attention model is implementation of Show, Attend, and Tell paper

In the implementation I get help from this repo

Pretrained Embeddings

I have downloaded the Fasttext trained embeddings for English

In embedding layer word2vec.py transformer is called for every word in the dictionary.

Obtained vectors create the weights of embedding layer.

Teacher Forcing

In dataset.py you can see how the target captions are created.

# Target captions are one step forward of train captions
target_captions = np.roll(train_captions, -1)

Target captions are one step forward of training captions.

Thus network tries to predict next word from previous word given for training.

Architecture

Here is feature extraction model. 'x_i' represents VGG16 output

Here is captioning architecture

Running

You can run in train, sample, and test modes e.g

$ python run.py train

Results

Our favorite wrong caption :)

Interestingly one of the actual captions below includes a caption related with soccer