Deploy tensorflow inception model as web service running on gunicorn/flask instead of grpc. Most of the stuffs are based on caffe web demo.
Assume that you have installed tensorflow, flask, gunicorn. There may be other python module dependencies, pip install them.
- git clone https://github.com/hetaoaoao/tensorflow_web_deploy
- Put your inception model (named like model.ckpt-5000 or tensorflow-serving exported 'export' file) and its meta file to "data" directory, rename them to model and model.meta. Also modify the synset file accordingly.
I already put a model.meta and a synset file there which is for the 5 flowers demo, trained model file is too large to upload to github. btw: the 5000 steps model get a 100% success rate:-)
Succesfully loaded model from /mnt/data1/tf/flowers5/models/model.ckpt-5000 at step=5000. 2016-07-21 16:16:04.621380: starting evaluation on (validation). 2016-07-21 16:16:09.724054: [20 batches out of 32] (125.4 examples/sec; 0.255sec/batch) 2016-07-21 16:16:11.959930: precision @ 1 = 1.0000 recall @ 5 = 1.0000 [1024 examples]
- Here we go, just 'sh run.sh', tail engine.log to check if something happens. If it works well, you should see something like.
INFO:PyClassification:classify_image cost 0.47 secs 2016-07-21 17:30:37 [4374] [INFO] sample testing complete roses sunflowers dandelio
when starting, it will try to classify the data/sample.jpg which is a rose picture.
- Open your browser, navigate to "http://localhost:5001", the page should be simple enough to understand.
- Use as webservice. Modify the interface for your need.
curl --request POST --data-binary "@sample.jpg" http://localhost:5001