Joint-GAN

Tensorflow implementation for reproducing results in Joint GAN. Implemented based on StackGAN. Many thanks for sharing the code.

Dependencies

• python 2.7
• TensorFlow 1.0.0
• prettytensor
• progressbar
• python-dateutil
• easydict
• pandas
• torchfile

Data

1. Download the preprocessed char-CNN-RNN text embeddings for birds and save them to Data/
2. Download the birds image data and extract to Data/birds/
3. Preprocess images: python misc/preprocess_birds.py

Pretrained Model

Download the pretrained LSTM decoder for bird and unzip all files to pretrain/

Training

Train a Joint GAN model on the CUB dataset using the preprocessed data for birds: python Main.py

Results

Generated results can be find in ckt_logs/birds/

• fake_images.jpg: generated images from noise
• gen_fake_sentences.txt: conditionally generated sentences based on fake_images.jpg
• fake_sentences.txt: generated sentences from noise
• gen_fake_images.jpg: conditionally generated images based on fake_sentences.txt

Images in the very left column of each file are the sample real images. The rest 16 images are paired with the first 16 sentences in the corresponding text file.