Skip to content

Latest commit

 

History

History
118 lines (79 loc) · 4.35 KB

README.md

File metadata and controls

118 lines (79 loc) · 4.35 KB
Raw

NotMinion

This APP is inspired by Not Hotdog of Jian Yang's idea in Silicon Valley, and it is just another version of it: Not Minion

Requirements

Introduction

It first retrains TensorFlow Inception v3's last layer to learn how a Minion image looks like, makes a model, then applies this model into TensorFlow's example iOS App to recognize a Minion.

If you want to build your own Not something from scratch, you should read from Step 1. Otherwise, you can jump to Step 5.

Step 1

First you have to install all tools listed in Requirements. Please refer to their websites and GitHub.

Step 2

Prepare for image sets.
TensorFlow Inception v3 is used to classify objects, so you have to divide images into 2 categories: Minions and Not Minions. And the root image directory should be organized as below:

└── Minion Images           <---- DIRECTORY_TO_YOUR_IMAGES
    ├── Minions
    └── Not Minions         <---- DIRECTORY_TO_IMAGE_DOWNLOAD

  1. You can download or scrape many Minion images from Internet, and put them into Minions.

  2. In Image Crawler there is a random image downloader for Not Minions. Usage:

python3 /Image Crawler/random_imgur.py -i 200 -o DIRECTORY_TO_IMAGE_DOWNLOAD
python2 /Image Crawler/delete_not_jpg.py -o DIRECTORY_TO_IMAGE_DOWNLOAD

Attention

  • random_imgur.py uses Python3, and delete_not_jpg.py uses Python2.
  • The DIRECTORY_TO_IMAGE_DOWNLOAD in two commands should be the same.

Step 3

Once you have installed TensorFlow, you can begin to retrain it to learn your something. In this App, it is Minion.

It means the image directory should contains 2 sub directories, each of them contains the corresponding images. (You have done this in Step 2)

Then execute the following commands in terminal:

cd ROOT_DIRECTORY_OF_TENSORFLOW
bazel build tensorflow/examples/image_retraining:retrain
bazel-bin/tensorflow/examples/image_retraining/retrain --image_dir DIRECTORY_TO_YOUR_IMAGES

This shall take a while, based on the performance of your computer.

Step 4

Now you have generated the output_graph.pb file in /tmp by default. Now in terminal, please enter

cd ROOT_DIRECTORY_OF_TENSORFLOW
bazel build tensorflow/examples/label_image:label_image && \
bazel-bin/tensorflow/examples/label_image/label_image \
--graph=/tmp/output_graph.pb --labels=/tmp/output_labels.txt \
--output_layer=final_result \
--input_layer=Mul \
--image=PATH_TO_YOUR_TEST_IMAGE

and witness Miracle!

A minion image:

2017-06-12 14:32:09.507037: I tensorflow/examples/label_image/main.cc:251] minions (1): 0.999147
2017-06-12 14:32:09.507057: I tensorflow/examples/label_image/main.cc:251] not minions (0): 0.000853157

Not a minion image:

2017-06-12 14:31:43.194735: I tensorflow/examples/label_image/main.cc:251] not minions (0): 0.890092
2017-06-12 14:31:43.194766: I tensorflow/examples/label_image/main.cc:251] minions (1): 0.109908

The float number at the end of each line represents the likelihood that this image is (not) a minion.

Step 5

If you just want to run this app on your iPhone, you can start from this step. You need to have TensorFlow prepared on your Mac.

  1. Clone this repo.

  2. Copy the /Not Minion TensorFlow Example directory (which is an example iOS App provided by TensorFlow with our own model applied) into ROOT_DIRECTORY_OF_TENSORFLOW/tensorflow/tensorflow/contrib/ios_examples/.

  3. Download the trained Minion model from BaiduDisk and copy it to ../Not Minion TensorFlow Example/data

    or

    Copy your own output_labels.txt and output_graph.pb to ../Not Minion TensorFlow Example/data and rename them as imagenet_comp_graph_label_strings.txt and tensorflow_inception_graph.pb.

  4. Open Not Minion TF.xcodeproj and run!

    Maybe you need to change Developer Certificates.

Preview

               

TODO

Transform Inception v3 model to Core ML format and apply it with CoreML Framework in iOS 11.

Created on 2017-07-05