Api for Image Recognition Service

Easy way to deploy multiple models with deep learning frameworks for your products.

深度学习模型多模型线上部署API. 可根据需要快速部署tensorflow, Keras模型(flask + gevent + gunicorn). 有更高级的需求可以使用tensorflow serving

My test environment:

• Ubuntu 16.04
• python3

Installation

python setup.py build
python setup.py install

Usage

from img_recog_api.model_creator import ImageModelSingle

config = 'path/to/config'

# image recognition
# create and load models
Model_s = ImageModelSingle(config)
# load image to api support format
image = your_load_image_func(image)
# predict image
results_s = Model_s.predict(image, top=4)

configs

Contains all config files.

• config.yaml

model_type: SingleTf
model_file: /path/to/model.pb
label_file: /path/to/labels.txt
# gpu_num could be set as any int. for cpu using gpu_num: ""
gpu_num: 0
input_tensor: input tensor name of tensorflow model(eg: Placeholder:0), if None, find it with pattern
output_tensor: output tensor name of tensorflow model(eg: Prediction:0), if None, find it with pattern

Outputs

ImageModelSingle()

model_type=SingleTf

**result = model.predict(image, top)**

{
  "recognitionList": [
    {
      "className": "label1",
      "confidence": 0.2440231442451477
    }
  ]
}

Add your own model

Just over write the sub_models.py and add new class for your model like the example SingleTf.

Deploy your models

In this demo, I deployed a serving model with python using flask + gevent + gunicorn.

Is easy to deploy multiple models in this way. But tensorflow serving maybe a good choose for long-term use.

How to use it:

• cd to example folder.
• add your prediction scripts in sub_model.py and install the api.
• change config.yaml according to your models.
• add config information in gunicorn.config. For detail information pleae view gunicorn setting document
• change image_recognition_all.py.

Run the server:

sh ./example_run.sh

Test it:

python example/service_test.py

TODO

• remove the warm-up part (If anyone comes up a good idea to speed up the first test, feel free to contact me :) )