im2txt_demo

im2txt + pretrained models + docker = easy tryout

This repo contains a copy of im2txt (https://github.com/tensorflow/models/tree/master/im2txt), a suitable docker image, pretrained models (tensorflow/models#466) and some trivial shell scripts for convenience.

Limitations:

• The command "git lfs pull" might respond with errors, due to the limited GitHub data quota:

batch response: This repository is over its data quota. Purchase more data packs to restore access. 

In this case you can

• wait until next month and hope you can perform "git lfs pull" before the data limit is reached OR
• use the following links to the pretrained models and download them manually: https://drive.google.com/file/d/0Bw6m_66JSYLlRFVKQ2tGcUJaWjA/view
• Copy the file model.ckpt-2000000 to the folder im2txt_pretrained. It should have a size of 143 mb.

The tensorflow version is fixed to an old one (0.12). Feel free to update and create a pull request! Thanks in Advance for your help!

Getting started:

Install docker (skip if already installed)

https://docs.docker.com/engine/installation/

Install git lfs (skip if already installed)

https://help.github.com/articles/installing-git-large-file-storage/

Clone this repo and pull large files (pretrained models):

git clone https://github.com/siavash9000/im2txt_demo.git
cd im2txt_demo
git lfs pull

build docker image

cd im2txt_demo
make build

start container

make run

run build_inference to prepare inference step (must be performed only once)

./build_inference.sh

get caption for image

./process_image.sh imgs/bikes.jpg