README.md

Experiments on community model zoos

Tensorflow.js

Postman

• express running on Node.js
• sharp(libvips)
• @tensorflow/tfjs-node

Running the server

$ cd tfjs-server
$ npm install

$ npm start
# Server listening on 3030

Running from Docker

$ cd tfjs-server
$ docker build -t incubating/tfjs-server:latest -f Dockerfile .
$ docker run -t --rm -p 3030:3030 "incubating/tfjs-server" &
# Server listening on 3030

Deploy to Elastic Beanstalk

$ brew install awsebcli
$ eb init
$ eb create
# t3.small, t2.small

$ eb deploy

Conventions

• GET route using ìmage_url query string or POST using multipart/form-data using the image field
• GET route using message query string or POST using message field in the JSON body
• postman/ contains a list of examples to try with the Postman application

tensorflow/toxicity

Annotations conventions

$ curl -s 'localhost:3030/v2/toxicity?text=I%20love%20seafood' | jq .
[
    {
        "label": "identity_attack",
        "results": [
            {
                "probabilities": {
                    "0": 0.9999690055847168,
                    "1": 0.000030998882721178234
                },
                "match": false
            }
        ]
    },
    {
        "label": "insult",
        "results": [
            {
                "probabilities": {
                    "0": 0.9998032450675964,
                    "1": 0.00019675538351293653
                },
                "match": false
            }
        ]
    },
    {
        "label": "obscene",
        "results": [
            {
                "probabilities": {
                    "0": 0.9999403953552246,
                    "1": 0.000059632253396557644
                },
                "match": false
            }
        ]
    },
    {
        "label": "severe_toxicity",
        "results": [
            {
                "probabilities": {
                    "0": 1,
                    "1": 2.4753488148121505e-8
                },
                "match": false
            }
        ]
    },
    {
        "label": "sexual_explicit",
        "results": [
            {
                "probabilities": {
                    "0": 0.9999597072601318,
                    "1": 0.00004024286681669764
                },
                "match": false
            }
        ]
    },
    {
        "label": "threat",
        "results": [
            {
                "probabilities": {
                    "0": 0.9999363422393799,
                    "1": 0.00006362887506838888
                },
                "match": false
            }
        ]
    },
    {
        "label": "toxicity",
        "results": [
            {
                "probabilities": {
                    "0": 0.9995335340499878,
                    "1": 0.00046650166041217744
                },
                "match": false
            }
        ]
    }
]

tensorflow/mobilenet

Classes it can predict: imagenet classes.

$ curl -X POST localhost:3030/v2/mobilenet
[
    {
        "className": "teddy, teddy bear",
        "probability": 0.9991821050643921
    },
    {
        "className": "hamper",
        "probability": 0.0001980842207558453
    },
    {
        "className": "toyshop",
        "probability": 0.00019572325982153416
    }
]

tensorflow/cocossd

$ curl -X POST localhost:3030/v2/cocossd
[
    {
        "bbox": [
            13.42470794916153,
            -2.2287483513355255,
            257.27201026678085,
            252.97450867295265
        ],
        "class": "teddy bear",
        "score": 0.9690217971801758
    }
]

Known issues

• node-gyp errors on MacOS: https://github.com/nodejs/node-gyp/blob/master/macOS_Catalina.md

References

Models

• Tensorflow / Toxicity classifier
• Tensorflow / Coco SSD
• Tensorflow / Mobilenet
• Tensorflow / USE
• naptha/tesseract.js
• lokesh/color-thief
• justadudewhohacks/face-api.js

To be evaluated

• Tensorflow / Posenet
• Tensorflow / BodyPix
• justadudewhohacks/opencv4nodejs