An API interface to some Caffe models. We use Flask to expose Caffe models previously trained.

We have used the Imagenet python wrapper from the Caffe repository and have implemented (based on that) a Lenet wrapper, in a local module _caffe. The image preprocess is done by the wrappers, but for the Lenet we must provide a centered image of a number, to match the structure of the images in the MNIST dataset used for training it.

You just need to set the path to your trained models in the file setup.py.

Run the script and wait for the models to load. After all models are loaded the server will be working on port 5000, the models being available in /lenet and /imagenet, respectively. You can POST a image to a model using, for example, the following command:

curl --form "image=@/path/to/image/image.jpg" [YOUR_IP]:5000/[MODEL_NAME]

The response is a JSON file containing the best class prediction, a list with the classes probabilities sorted by descending order and a list with the classes sorted by the propabiblities in descending order. Here is a response example:

{
  "best": 0, 
  "predictions": [
    0, 
    6, 
    9, 
    7, 
    8, 
    5, 
    2, 
    1, 
    3, 
    4
  ], 
  "probabilities": [
    0.9979206919670105, 
    0.0007992387982085347, 
    0.0005818080389872193, 
    0.00036282758810557425, 
    0.0001612447085790336, 
    0.00010736921103671193, 
    4.2144984035985544e-05, 
    1.3334250979823992e-05, 
    6.253382252907613e-06, 
    5.1289121074660216e-06
  ]
}