Use Flask framework and fine-tuning MobileNet weights from Keras to build a simple AI web application.
Core Function: Classify the input image to dog or cat..yeah, pretty simple function.
Python is not the mainstream option for web development, however, it would be much easier to embed the machine learning script into a app developed with Python. And Flask is a light-weight framework for beginners. I don't know anything about JavaScript, but Jinja2 provides simple (but ugly) solution for the front end. Therefore, I chose Flask.
As for Keras, it offers pretrained weights for different calssic models, very easy to do transfer learning. Since it's only a local project without GPU server, Mobilenet is a good option to run on CPU. Actually there are more efficient models, like ShuffleNet.
keras 2.0.8
tensorflow 1.2.1
flask 0.12.2
python 3.5.2
- git clone this repo
- set up the virtual environment
- go to the path of python.exe
set FLASK_APP=path
set FLASK_ENV=development
flask run
where the path is the absolute path of app.py
Note: Use export instead of set on Linux OS. There should be no space around =！
The app allows users to upload jpg or png images of cats or dogs and returns the classification and probability. It allows the user to provide feedback about the classification and use it to update the weights of MobileNet.
- Upload the file
Only support files with English name. If the name is in Chinese, it will show a hint.
- It returns the classification result and probability. You can also use feedback button to help immprove the model.
- Thanks for the feedback
│ app.py
│ repredict.py load weights to classify
│ TryMobile.py Update weights
├─static
│ │ cats_dog_mobileNet.h5 weights after transfer learning
│ │ files_cat.npy
│ │ files_dog.npy numpy array to save paths of newly labeled images
│ │ style.css
│ └─images
│ │ save unlabeled input images
│ ├─Cat
│ │ save images labeled as 'Cat' for updating weights
│ └─Dog
├─templates
│ show.html the interface to upload images
│ thanks.html
│ upload.html show the classification result and users can give feedbacks
└─uploads