text2image-benchmark

Performance comparison of existing GAN based Text To Image algorithms. (GAN-CLS, StackGAN, TAC-GAN)
Note: The codes in algorithms folder is brought from the respective author's repo, not written by me.

Bug fixed.

• StackGAN:

• modified requirements.txt to avoid environemnt conflict.

• used tensorflow 1.0.1 version instead of 0.12.0, to avoid tf.zeros_initializer() error.

• change the argument order of tf.concat to avoid type mismatch error. [see here]

• added folder/file creation code to automatically create Data/birds/example_captions.txt file. (please add example sentences to example_captions.txt file, otherwise there will be error.)

• GAN-CLS:

   (you may need to manually do this to install tensorflow 0.11.0 version using pip)
   export TF_BINARY_URL=https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow-0.11.0-cp27-none-linux_x86_64.whl
   sudo pip install --upgrade $TF_BINARY_URL

• used tensorflow 0.11.0 version instead of 0.12.0, to avoidi get_variable() error.

How to run?

• StackGAN

. venv/bin/activate
python demo/birds_skip_thought_demo.py --cfg demo/cfg/birds-skip-thought-demo.yml --gpu 0 --caption_path <your_text_sentences_path>