Python 3 and ResNet feature version by @milhidaka

Including caption generation demo on web browser using WebDNN.

• Python 3.6
• Chainer 2.0.0
• Cupy 1.0.0
• Pillow

Download caption_gen_resnet.model (45MB) and dataset_coco.pkl (28MB).

$ python src/generate_caption.py -s dataset_coco.pkl -m caption_gen_resnet.model -l image/list.txt -g 0

Options:

• -s, sentence: (required) sentence dataset file path.
• -m, --model: (required) trained model file path.
• -l, --list: (required) image path list file.
• -g, --gpu: (optional) GPU index. -1 means CPU.

$ python src/convert_webdnn.py --sentence dataset_coco.pkl --model caption_gen_resnet.model --example_image image/asakusa.jpg

Then start a HTTP server (python -m http.server) and go to http://localhost:8000/webdnn.

1. Download images (2014) from http://mscoco.org/dataset/#download and extract to some directory.
2. Download caption_datasets.zip from: http://cs.stanford.edu/people/karpathy/deepimagesent/
3. Extract downloaded zip file, and you'll get dataset_coco.json.

$ python src/convert_dataset.py dataset_coco.json dataset_coco.pkl

Parameters:

• sentence JSON file of dataset.
• output pkl file.

$ python src/extract_resnet_feat.py dataset_coco.json /path/to/coco/images resnet_feat.mat -g 0 -b 16

Options:

• sentence JSON file of dataset.
• Top-level directory containing images. Searches files recursively.
• output feature matrix file. (becomes about 1GB)
• -g, --gpu: (optional) GPU index. -1 means CPU.
• -b, --batchsize: (optional) batch size for extracting feature.

It will take several hours.

$ python src/train.py -g 0 -s dataset_coco.pkl -i resnet_feats.mat -o model/caption_gen

Options:

• -g, --gpu: (optional) GPU device index (default: -1).
• -s, --sentence: (required) sentence dataset file path.
• -i, --image: (required) image feature file path.
• -m, --model: (optional) input model file path without extension.
• -o, --output: (required) output model file path without extension.
• --iter: (optional) the number of iterations (default: 100).

image/asakusa.jpg
image/tree.jpg

MIT License