A high performance node.js RedisAI client

You can use this client to query RedisAI using plain Javascript, or in a type-safe manner using Typescript, since redisai-js comes with its own type definitions built-in.

Installation

Installation is done using the npm install command:

npm install --save redisai-js

Overview

Vanilla JS:

Example of AI.TENSORSET and AI.TENSORGET

var redis = require('redis');
var redisai = require('redisai-js');

(async () => {
    const nativeClient = redis.createClient();
    const aiclient = new redisai.Client(nativeClient);
    const tensorA = new redisai.Tensor(redisai.Dtype.float32, [1, 2], [3, 5]);
    const result = await aiclient.tensorset("tensorA", tensorA);

    // AI.TENSORSET result: OK
    console.log(`AI.TENSORSET result: ${result}`)

    const tensorGetReply = await aiclient.tensorget("tensorA");

    // AI.TENSORGET reply: datatype FLOAT shape [1,2] , data [3,5]
    console.log(`AI.TENSORGET reply: datatype ${tensorGetReply.dtype} shape [${tensorGetReply.shape}] , data [${tensorGetReply.data}]`);

    await aiclient.end();
})();

Example of AI.MODELSET and AI.MODELRUN

var redis = require('redis');
var redisai = require('redisai-js');
var fs = require("fs");

(async () => {
    const nativeClient = redis.createClient();
    const aiclient = new redisai.Client(nativeClient);
    const tensorA = new redisai.Tensor(redisai.Dtype.float32, [1, 2], [2, 3]);
    const tensorB = new redisai.Tensor(redisai.Dtype.float32, [1, 2], [3, 5]);
    const result_tA = await aiclient.tensorset("tA", tensorA);
    const result_tB = await aiclient.tensorset("tB", tensorB);

    const model_blob = fs.readFileSync("./test_data/graph.pb");
    // AI.TENSORSET tA result: OK
    console.log(`AI.TENSORSET tA result: ${result_tA}`)
    // AI.TENSORSET tB result: OK
    console.log(`AI.TENSORSET tB result: ${result_tB}`)

    const mymodel = new redisai.Model(redisai.Backend.TF, "CPU", ["a", "b"], ["c"], model_blob);

    const result_modelSet = await aiclient.modelset("mymodel", mymodel);

    // AI.MODELSET result: OK
    console.log(`AI.MODELSET result: ${result_modelSet}`)

    const result_modelRun = await aiclient.modelrun("mymodel", ["tA", "tB"], ["tC"]);

    console.log(`AI.MODELRUN result: ${result_modelRun}`)
    const tensorC = await aiclient.tensorget("tC");

    // AI.TENSORGET tC reply: datatype FLOAT shape [1,2] , data [6,15]
    console.log(`AI.TENSORGET tC reply: datatype ${tensorC.dtype} shape [${tensorC.shape}] , data [${tensorC.data}]`);

    await aiclient.end();
})();

Example of AI.SCRIPTSET and AI.SCRIPTRUN

(async () => {
    const nativeClient = redis.createClient();
    const aiclient = new redisai.Client(nativeClient);
    const tensorA = new redisai.Tensor(redisai.Dtype.float32, [1, 2], [2, 3]);
    const tensorB = new redisai.Tensor(redisai.Dtype.float32, [1, 2], [3, 5]);
    const script_str = 'def bar(a, b):\n    return a + b\n';

    const result_tA = await aiclient.tensorset("tA", tensorA);
    const result_tB = await aiclient.tensorset("tB", tensorB);
    // AI.TENSORSET tA result: OK
    console.log(`AI.TENSORSET tA result: ${result_tA}`)
    // AI.TENSORSET tB result: OK
    console.log(`AI.TENSORSET tB result: ${result_tB}`)

    const myscript = new redisai.Script("CPU", script_str);

    const result_scriptSet = await aiclient.scriptset("myscript", myscript);

    // AI.SCRIPTSET result: OK
    console.log(`AI.SCRIPTSET result: ${result_scriptSet}`)

    const result_scriptRun = await aiclient.scriptrun("myscript", "bar",["tA", "tB"], ["tD"]);

    console.log(`AI.SCRIPTRUN result: ${result_scriptRun}`)
    const tensorD = await aiclient.tensorget("tD");

    // AI.TENSORGET tD reply: datatype FLOAT shape [1,2] , data [5,8]
    console.log(`AI.TENSORGET tD reply: datatype ${tensorD.dtype} shape [${tensorD.shape}] , data [${tensorD.data}]`);

    await aiclient.end();
})();

Example of AI.DAGRUN enqueuing multiple SCRIPTRUN and MODELRUN commands

A common pattern is enqueuing multiple SCRIPTRUN and MODELRUN commands within a DAG. The following example uses ResNet-50,to classify images into 1000 object categories.

Given that our input tensor contains each color represented as a 8-bit integer and that neural networks usually work with floating-point tensors as their input we need to cast a tensor to floating-point and normalize the values of the pixels - for that we will use pre_process_4ch function.

To optimize the classification process we can use a post process script to return only the category position with the maximum classification - for that we will use post_process script.

Using the DAG capabilities we've removed the necessity of storing the intermediate tensors in the keyspace. You can even run the entire process without storing the output tensor, as follows:

var redis = require('redis');
var redisai = require('redisai-js');
var fs = require("fs");
var Jimp = require('jimp');

(async () => {
    const nativeClient = redis.createClient();
    const aiclient = new redisai.Client(nativeClient);
    const scriptFileStr = fs.readFileSync('./tests/test_data/imagenet/data_processing_script.txt').toString();
    const jsonLabels = fs.readFileSync('./tests/test_data/imagenet/imagenet_class_index.json');
    const labels = JSON.parse(jsonLabels);

    const dataProcessingScript = new redisai.Script('CPU', scriptFileStr);
    const resultScriptSet = await aiclient.scriptset('data_processing_script', dataProcessingScript);
    // AI.SCRIPTSET result: OK
    console.log(`AI.SCRIPTSET result: ${resultScriptSet}`)

    const modelBlob = fs.readFileSync('./tests/test_data/imagenet/resnet50.pb');
    const imagenetModel = new redisai.Model(redisai.Backend.TF, 'CPU', ['images'], ['output'], modelBlob);
    const resultModelSet = await aiclient.modelset('imagenet_model', imagenetModel);
    
    // AI.MODELSET result: OK
    console.log(`AI.MODELSET result: ${resultModelSet}`)

    const inputImage = await Jimp.read('./tests/test_data/imagenet/cat.jpg');
    const imageWidth = 224;
    const imageHeight = 224;
    const image = inputImage.cover(imageWidth, imageHeight);
    const tensor = new redisai.Tensor(redisai.Dtype.uint8, [imageWidth, imageHeight, 4], Buffer.from(image.bitmap.data));
    
    ///
    // Prepare the DAG enqueuing multiple SCRIPTRUN and MODELRUN commands
    const dag = new redisai.Dag();
    
    dag.tensorset('tensor-image', tensor);
    dag.scriptrun('data_processing_script', 'pre_process_4ch', ['tensor-image'], ['temp_key1']);
    dag.modelrun('imagenet_model', ['temp_key1'], ['temp_key2']);
    dag.scriptrun('data_processing_script', 'post_process', ['temp_key2'], ['classification']);
    dag.tensorget('classification');

    // Send the AI.DAGRUN command to RedisAI server
    const resultDagRun = await aiclient.dagrun_ro(null, dag);
    
    // The 5th element of the reply will be the `classification` tensor 
    const classTensor = resultDagRun[4];
    
    // Print the category in the position with the max classification
    const idx = classTensor.data[0];

    // 281 [ 'n02123045', 'tabby' ]
    console.log(idx, labels[idx.toString()]);

    await aiclient.end();
})();

Further examples

The RedisAI examples repo shows more advanced examples made using redisai-js under js_client folder.

Supported RedisAI Commands

Command	Recommended API and JSDoc
AI.TENSORSET	tensorset
AI.TENSORGET	tensorget
AI.MODELSET	modelset
AI.MODELGET	modelget
AI.MODELDEL	modeldet
AI.MODELRUN	modelrun
AI._MODELSCAN	N/A
AI.SCRIPTSET	scriptset
AI.SCRIPTGET	scriptget
AI.SCRIPTDEL	scriptdel
AI.SCRIPTRUN	scriptrun
AI._SCRIPTSCAN	N/A
AI.DAGRUN	dagrun
AI.DAGRUN_RO	dagrun_ro
AI.INFO	info and infoResetStat (for resetting stats)
AI.CONFIG *	configLoadBackend and configBackendsPath

Running tests

A simple test suite is provided, and can be run with:

$ npm test

The tests expect a Redis server with the RedisAI module loaded to be available at localhost:6379

License

redisai-js is distributed under the BSD3 license - see LICENSE