Stylenet Pytorch
Pytorch Implementation of Stylenet
This implementation is based on Pytorch 1.0
Setup
To install required modules run command below.
pipenv install
pipenv shell
Prepare dataset
Preprocess tokens for pytorch data loader.
python stylenet/text_preprocessor.py \
-normal_path <path to non-style caption> \
-style_path <path to style specific caption> \
-save_path <path to save name of vocab file> \
-max_vocab_size 0 # set 0 when using all vocaburaly. \
# -train_embed_matrix # set this if you want to train pre-trained embedding. \
-dim_size 512 \
-window_size 5
or just put Flickr8K dataset and FlickrStyle to 'data/train_data' and type it.
sh sh/preprocess.sh
Train
configuration for training and inference is based on data/configs/default.yml
python stylenet/train.py -config_path data/configs/default.yml
inference
python stylenet/generate.py \
-img_dir <path to image directory>/ # it can be /sample_imgs/ \
-text_preprocessor_path <path to save name of vocab file created by text_preprocessor.py> \
-checkpoint_path <path to saved checkpoint> \
-mode style # or default \
-decoder beam # beam or greedy \
-beam_width 5 \
-gen_max_len 50 \
-config_path data/configs/default.yml \