You're Invited:Meet the Socket Team at BlackHat and DEF CON in Las Vegas, Aug 4-6.RSVP →
Socket
Book a DemoInstallSign in
Socket
Book a DemoInstallSign in

menoh

Package Overview
Dependencies
Maintainers
1
Versions
7
Alerts
File Explorer

Advanced tools

License
Socket logo

Install Socket

Detect and block malicious and high-risk dependencies

Install

menoh - npm Package Compare versions

Comparing version

1.0.1
to
1.1.0

98

example/example_mnist.js
'use strict';
const jimp = require("jimp");
const assert = require('assert');
const fs = require('fs');
const ndarray = require('ndarray');
const dtype = require('dtype');
const _ = require('lodash');

@@ -32,6 +33,6 @@ const menoh = require('..'); // This menoh module

var outp = [];
for (var i = 0; i < a.length; i++) {
for (var i = 0; i < a.size; i++) {
outp.push(i); // add index to output array
if (outp.length > k) {
outp.sort((l, r) => { return a[r] - a[l]; });
outp.sort((l, r) => { return a.get(r) - a.get(l); });
outp.pop();

@@ -47,68 +48,67 @@ }

loadInputImages()
.then((imageList) => {
const data = [];
imageList.forEach((image, batchIdx) => {
// All the input images are already croped and resized to 28 x 28.
assert.equal(image.bitmap.width, 28);
assert.equal(image.bitmap.height, 28);
/*
*/
// Convert bitmap to an array. (Use R channel only - already in greyscale)
const numPixels = image.bitmap.width * image.bitmap.height;
const batchOffset = batchIdx * numPixels;
image.scan(0, 0, image.bitmap.width, image.bitmap.height, function (x, y, idx) {
const dataIdx = y * image.bitmap.width + x + batchOffset;
data[dataIdx] = this.bitmap.data[idx]; // R channel
});
});
// Load ONNX file
return menoh.create('../test/data/mnist/mnist.onnx')
.then((builder) => {
const batchSize = INPUT_IMAGE_LIST.length;
// Load ONNX file
return menoh.create('../test/data/mnist/mnist.onnx')
.then((builder) => {
const batchSize = imageList.length;
// Add input data
builder.addInput(MNIST_IN_NAME, [
batchSize, // 10 images in the data
1, // number of channels
28, // height
28 // width
]);
// Add input data
builder.addInput(MNIST_IN_NAME, [
batchSize, // 10 images in the data
1, // number of channels
28, // height
28 // width
]);
// Add output
builder.addOutput(MNIST_OUT_NAME);
// Add output
builder.addOutput(MNIST_OUT_NAME);
// Build a new Model
const model = builder.buildModel({
backendName: 'mkldnn'
})
// Build a new Model
const model = builder.buildModel({
backendName: 'mkldnn'
})
// Create a view for input buffer using ndarray.
const iData = (function () {
const prof = model.getProfile(MNIST_IN_NAME);
return ndarray(new (dtype(prof.dtype))(prof.buf.buffer), prof.dims);
})();
// Set input data
model.setInputData(MNIST_IN_NAME, data);
// Create a view for output buffer using ndarray.
const oData = (function () {
const prof = model.getProfile(MNIST_OUT_NAME);
return ndarray(new (dtype(prof.dtype))(prof.buf.buffer), prof.dims);
})();
return loadInputImages()
.then((imageList) => {
imageList.forEach((image, batchIdx) => {
// All the input images are already croped and resized to 28 x 28.
// Now, copy the image data into to the input buffer in NCHW format.
image.scan(0, 0, image.bitmap.width, image.bitmap.height, (x, y, idx) => {
const val = image.bitmap.data[idx];
iData.set(batchIdx, 0, y, x, val);
});
});
// Run the model
return model.run()
.then(() => {
const out = model.getOutput(MNIST_OUT_NAME);
// just to be sure
assert.equal(out.dims[0] * out.dims[1], out.data.length);
assert.equal(out.dims[0], batchSize); // only applies to this example
// Print the results.
out.data = _.chunk(out.data, out.dims[1]); // reshaped
for (let bi = 0; bi < batchSize; ++bi) {
console.log('### Result for %s', INPUT_IMAGE_LIST[bi]);
const topK = findIndicesOfTopK(out.data[bi], 1);
const topK = findIndicesOfTopK(oData.pick(bi, null), 1);
topK.forEach((i) => {
console.log('[%d] %f %s', i, out.data[bi][i], categoryList[i]);
console.log('[%d] %f %s', i, oData.get(bi, i), categoryList[i]);
});
}
// Happily done!
});
});
})
.catch((err) => {
console.log('Error:', err);
});

117

example/example_vgg16.js

@@ -6,2 +6,4 @@ 'use strict';

const fs = require('fs');
const ndarray = require('ndarray');
const dtype = require('dtype');
const _ = require('lodash');

@@ -51,6 +53,6 @@ const menoh = require('..'); // This menoh module

var outp = [];
for (var i = 0; i < a.length; i++) {
for (var i = 0; i < a.size; i++) {
outp.push(i); // add index to output array
if (outp.length > k) {
outp.sort((l, r) => { return a[r] - a[l]; });
outp.sort((l, r) => { return a.get(r) - a.get(l); });
outp.pop();

@@ -66,79 +68,80 @@ }

loadInputImages()
.then((imageList) => {
const data = [];
imageList.forEach((image, batchIdx) => {
// Crop the input image to a square shape.
cropToSquare(image);
// Load ONNX file
return menoh.create('../test/data/vgg16/VGG16.onnx')
.then((builder) => {
const batchSize = INPUT_IMAGE_LIST.length;
// Resize it to 224 x 224.
image.resize(224, 224);
// Add input
builder.addInput(CONV1_1_IN_NAME, [
batchSize, // 2 images in the data
3, // number of channels
224, // height
224 // width
]);
// Convert bitmap to an array.
const numPixels = image.bitmap.width * image.bitmap.height;
const batchOffset = batchIdx * numPixels * 3;
image.scan(0, 0, image.bitmap.width, image.bitmap.height, function (x, y, idx) {
for (let c = 0; c < 3; ++c) {
const dataIdx = c * numPixels + y * image.bitmap.width + x + batchOffset;
data[dataIdx] = this.bitmap.data[idx + c];
}
});
});
// Add output
builder.addOutput(FC6_OUT_NAME);
builder.addOutput(SOFTMAX_OUT_NAME);
// Load ONNX file
return menoh.create('../test/data/vgg16/VGG16.onnx')
.then((builder) => {
const batchSize = imageList.length;
// Build a new Model
const model = builder.buildModel({
backendName: 'mkldnn'
})
// Add input
builder.addInput(CONV1_1_IN_NAME, [
batchSize, // 2 images in the data
3, // number of channels
224, // height
224 // width
]);
// Create a view for input buffer using ndarray.
const iData = (function () {
const prof = model.getProfile(CONV1_1_IN_NAME);
return ndarray(new (dtype(prof.dtype))(prof.buf.buffer), prof.dims);
})();
// Add output
builder.addOutput(FC6_OUT_NAME);
builder.addOutput(SOFTMAX_OUT_NAME);
// Create a view for each output buffers using ndarray.
const oDataFc6 = (function () {
const prof = model.getProfile(FC6_OUT_NAME);
return ndarray(new (dtype(prof.dtype))(prof.buf.buffer), prof.dims);
})();
const oDataSmx = (function () {
const prof = model.getProfile(SOFTMAX_OUT_NAME);
return ndarray(new (dtype(prof.dtype))(prof.buf.buffer), prof.dims);
})();
// Build a new Model
const model = builder.buildModel({
backendName: 'mkldnn'
})
return loadInputImages()
.then((imageList) => {
const data = [];
imageList.forEach((image, batchIdx) => {
// Crop the input image to a square shape.
cropToSquare(image);
// Set input data
model.setInputData(CONV1_1_IN_NAME, data);
// Resize it to 224 x 224.
image.resize(224, 224);
// Now, copy the image data into to the input buffer in NCHW format.
image.scan(0, 0, image.bitmap.width, image.bitmap.height, (x, y, idx) => {
for (let c = 0; c < 3; ++c) {
const val = image.bitmap.data[idx + c];
iData.set(batchIdx, c, y, x, val);
}
});
});
// Run the model
return model.run()
.then(() => {
const out1 = model.getOutput(FC6_OUT_NAME);
const out2 = model.getOutput(SOFTMAX_OUT_NAME);
// just to be sure
assert.equal(out1.dims[0] * out1.dims[1], out1.data.length);
assert.equal(out2.dims[0] * out2.dims[1], out2.data.length);
assert.equal(out1.dims[0], batchSize); // only applies to this example
assert.equal(out2.dims[0], batchSize); // only applies to this example
// Print the results.
out1.data = _.chunk(out1.data, out1.dims[1]); // reshaped
out2.data = _.chunk(out2.data, out2.dims[1]); // reshaped
for (let bi = 0; bi < batchSize; ++bi) {
console.log('### Result for %s', INPUT_IMAGE_LIST[bi]);
console.log('fc6 out: %s ...', out1.data[bi].slice(0, 5).join(' '));
const fc6 = oDataFc6.pick(bi, null);
console.log('fc6 out: %s ...', [0, 1, 2].map((i) => fc6.get(i)).join(' '));
const topK = findIndicesOfTopK(out2.data[bi], 5);
const topK = findIndicesOfTopK(oDataSmx.pick(bi, null), 5);
console.log('Top 5 categories are:');
topK.forEach((i) => {
console.log('[%d] %f %s', i, out2.data[bi][i], categoryList[i]);
console.log('[%d] %f %s', i, oDataSmx.get(bi, i), categoryList[i]);
});
}
// Happily done!
});
});
})
.catch((err) => {
console.log('Error:', err);
});

9

package.json
{
"name": "menoh",
"version": "1.0.1",
"version": "1.1.0",
"description": "NodeJS binding for Menoh DNN inference library.",

@@ -18,3 +18,4 @@ "main": "index.js",

"dnn",
"mkl-dnn"
"mkl-dnn",
"chainer"
],

@@ -29,4 +30,6 @@ "author": "enobufs",

"devDependencies": {
"dtype": "^2.0.0",
"jimp": "^0.2.28",
"mocha": "^5.0.2"
"mocha": "^5.0.2",
"ndarray": "^1.0.18"
},

@@ -33,0 +36,0 @@ "engines": {

28

README.md
# menoh
NodeJS binding for Menoh DNN inference library.
## Features
* Fast DNN inference on Intel CPU.
* Support standard [ONNX](http://onnx.ai/) format.
* Easy to use.
## Requirements
* MKL-DNN library [v0.14](https://github.com/intel/mkl-dnn/tree/v0.14) or later.
* ProtocolBuffers (Only tested with v3.5.1)
* [Menoh(C/C++) library](https://github.com/pfnet-research/menoh) v1.x (Only tested with v1.0.2)
* ProtocolBuffers (Tested with v3.5.1)
* [Menoh(C/C++) library](https://github.com/pfnet-research/menoh) v1.x (Tested with v1.0.2)
* NodeJS v6 or greater

@@ -25,3 +30,3 @@

For linux, you may need add `/usr/local/lib` to LD_LIBRARY_PATH depending on your linux distrubtion.
For linux, you may need to add `/usr/local/lib` to LD_LIBRARY_PATH depending on your linux distrubtion.
```sh

@@ -139,5 +144,11 @@ export LD_LIBRARY_PATH=/usr/local/lib

### Model methods
#### model.setInputData(input_var_name{string}, data{array})
Sets input data for the give input name.
#### model.getProfile(var_name{string}) => {object}
Returns an object for the given name.
The object has following properties:
* dims {array}: Output buffer dimensions. (e.g. [1, 3, 244, 244])
* buf {Buffer}: Output buffer attached to the variable.
* dtype {string}: Data type.
> Current revision supports only one data type, "float32".
#### model.run(cb) => {Promise}

@@ -148,3 +159,10 @@ Run inference. It returns promise if `cb` is not provided. The actual inference takes place

#### model.setInputData(input_var_name{string}, data{array})
> DEPREACATED. Use model.getProfile() instead.
Sets input data for the give input name.
#### model.getOutput(output_var_name) => {object}
> DEPREACATED. Use model.getProfile() instead.
Returns output object generated during `model.run()` for the given output name.

@@ -151,0 +169,0 @@ The output object has following properties:

262

test/tests.js

@@ -5,2 +5,4 @@ 'use strict';

const jimp = require("jimp");
const ndarray = require("ndarray");
const dtype = require("dtype");
const fs = require('fs');

@@ -48,11 +50,26 @@ const _ = require('lodash');

function createBufferView(model, name) {
const prof = model.getProfile(name);
return ndarray(new (dtype(prof.dtype))(prof.buf.buffer), prof.dims);
}
// Find the indexes of the k largest values.
function findIndicesOfTopK(a, k) {
var outp = [];
for (var i = 0; i < a.length; i++) {
outp.push(i); // add index to output array
if (outp.length > k) {
outp.sort((l, r) => { return a[r] - a[l]; });
outp.pop();
if (Array.isArray(a)) {
for (var i = 0; i < a.length; i++) {
outp.push(i); // add index to output array
if (outp.length > k) {
outp.sort((l, r) => { return a[r] - a[l]; });
outp.pop();
}
}
} else {
for (var i = 0; i < a.size; i++) {
outp.push(i); // add index to output array
if (outp.length > k) {
outp.sort((l, r) => { return a.get(r) - a.get(l); });
outp.pop();
}
}
}

@@ -62,16 +79,25 @@ return outp;

function validateOutput(model, batchSize) {
const out = model.getOutput(MNIST_OUT_NAME);
// sanity check
assert.equal(out.dims[0] * out.dims[1], out.data.length);
assert.equal(out.dims[0], batchSize);
function validateOutput(output, batchSize) {
if (Array.isArray(output.data)) {
// sanity check
assert.equal(output.dims[0] * output.dims[1], output.data.length);
assert.equal(output.dims[0], batchSize);
// Evaluate results.
out.data = _.chunk(out.data, out.dims[1]); // reshaped
for (let bi = 0; bi < batchSize; ++bi) {
const topK = findIndicesOfTopK(out.data[bi], 1);
topK.forEach((idx) => {
assert.equal(idx, bi);
});
// Evaluate results.
output.data = _.chunk(output.data, output.dims[1]); // reshaped
for (let bi = 0; bi < batchSize; ++bi) {
const topK = findIndicesOfTopK(output.data[bi], 1);
topK.forEach((idx) => {
assert.equal(idx, bi);
});
}
} else {
for (let bi = 0; bi < batchSize; ++bi) {
const topK = findIndicesOfTopK(output.pick(bi, null), 1);
topK.forEach((idx) => {
assert.equal(idx, bi);
});
}
}
}

@@ -113,8 +139,19 @@

model.setInputData(MNIST_IN_NAME, data);
const iv = createBufferView(model, MNIST_IN_NAME);
const ov = createBufferView(model, MNIST_OUT_NAME);
assert.equal(iv.size, data.length);
assert.deepEqual(iv.shape, [batchSize, 1, 28, 28]);
assert.equal(ov.size, 100);
assert.deepEqual(ov.shape, [batchSize, 10]);
// Write input data to input view.
data.forEach((v, i) => {
iv.data[i] = v;
});
// Run the model
model.run((err) => {
assert.ifError(err);
validateOutput(model, batchSize);
validateOutput(ov, batchSize);
done();

@@ -125,3 +162,3 @@ });

it('Succeed with promise', function () {
it('Succeed using views', function () {
// Load ONNX file

@@ -140,8 +177,19 @@ return menoh.create(ONNX_FILE_PATH)

model.setInputData(MNIST_IN_NAME, data);
const iv = createBufferView(model, MNIST_IN_NAME);
const ov = createBufferView(model, MNIST_OUT_NAME);
assert.equal(iv.size, data.length);
assert.deepEqual(iv.shape, [batchSize, 1, 28, 28]);
assert.equal(ov.size, 100);
assert.deepEqual(ov.shape, [batchSize, 10]);
// Write input data to input view.
data.forEach((v, i) => {
iv.data[i] = v;
});
// Run the model
return model.run()
.then(() => {
validateOutput(model, batchSize);
validateOutput(ov, batchSize);
});

@@ -165,16 +213,21 @@ });

model.setInputData(MNIST_IN_NAME, data);
const iv = createBufferView(model, MNIST_IN_NAME);
const ov = createBufferView(model, MNIST_OUT_NAME);
data.forEach((v, i) => {
iv.data[i] = v;
});
// Run the model
return model.run() // 1st run
.then(() => {
validateOutput(model, batchSize);
validateOutput(ov, batchSize);
return model.run(); // 2nd run
})
.then(() => {
validateOutput(model, batchSize);
validateOutput(ov, batchSize);
return model.run(); // 3rd run
})
.then(() => {
validateOutput(model, batchSize);
validateOutput(ov, batchSize);
});

@@ -197,4 +250,6 @@ });

})
model1.setInputData(MNIST_IN_NAME, data);
const iv1 = createBufferView(model1, MNIST_IN_NAME);
const ov1 = createBufferView(model1, MNIST_OUT_NAME);
// create model2

@@ -204,10 +259,15 @@ const model2 = builder.buildModel({

})
model2.setInputData(MNIST_IN_NAME, data);
const iv2 = createBufferView(model2, MNIST_IN_NAME);
const ov2 = createBufferView(model2, MNIST_OUT_NAME);
data.forEach((v, i) => {
iv1.data[i] = v;
iv2.data[i] = v;
});
// Run these models concurrently
return Promise.all([ model1.run(), model2.run() ])
.then(() => {
[model1, model2].forEach((model) => {
validateOutput(model, batchSize);
});
validateOutput(ov1, batchSize);
validateOutput(ov2, batchSize);
});

@@ -431,4 +491,4 @@ });

describe('#setInputData tests', function () {
it('should throw with invalid input data', function () {
describe('#run tests', function () {
it('should throw with invalid arg 1', function () {
return menoh.create(ONNX_FILE_PATH)

@@ -442,11 +502,17 @@ .then((builder) => {

model.setInputData('bad_input_name', data); // should throw
const iv = createBufferView(model, MNIST_IN_NAME);
// Write input data to input view.
data.forEach((v, i) => {
iv.data[i] = v;
});
model.run('bad')
})
.then(assert.fail, (err) => {
assert.ok(err instanceof Error);
assert.ok(err.message.includes('variable not found'));
assert.ok(err.message.includes('arg 1'));
});
});
it('should throw if input data is too short', function () {
it('second run() on the same model should fail', function () {
return menoh.create(ONNX_FILE_PATH)

@@ -460,12 +526,75 @@ .then((builder) => {

const tooShort = [0, 1, 2];
model.setInputData(MNIST_IN_NAME, tooShort); // should throw
const iv = createBufferView(model, MNIST_IN_NAME);
const ov = createBufferView(model, MNIST_OUT_NAME);
// Write input data to input view.
data.forEach((v, i) => {
iv.data[i] = v;
});
let err1 = null;
let err2 = null;
return Promise.all([
model.run().catch((err) => {
err1 = err;
}),
model.run().catch((err) => {
err2 = err;
}),
])
.then(() => {
assert.ok(!err1);
assert.ok(err2 instanceof Error);
assert.ok(err2.message.includes('in progress'));
validateOutput(ov, batchSize);
});
})
.then(assert.fail, (err) => {
assert.ok(err instanceof Error);
assert.ok(err.message.includes('too short'));
});
});
});
describe('Deprecated feature tests', function () {
let imageList;
let batchSize;
let data;
before(function () {
return loadInputImages(INPUT_IMAGE_LIST)
.then((_imageList) => {
imageList = _imageList;
batchSize = imageList.length;
data = preprocessImages(imageList);
});
});
it('Succeed with setInputData and getOutput', function () {
// Load ONNX file
return menoh.create(ONNX_FILE_PATH)
.then((builder) => {
const batchSize = imageList.length;
builder.addInput(MNIST_IN_NAME, [ batchSize, 1, 28, 28 ]);
builder.addOutput(MNIST_OUT_NAME);
// Make a new Model
const model = builder.buildModel({
backendName: 'mkldnn'
})
model.setInputData(MNIST_IN_NAME, data);
// Run the model
return model.run()
.then(() => {
const out = model.getOutput(MNIST_OUT_NAME);
validateOutput(out, batchSize);
});
});
});
it('should throw if input data is too long', function () {
describe('#setInputData failure tests', function () {
it('should throw with invalid input data', function () {
return menoh.create(ONNX_FILE_PATH)

@@ -479,14 +608,11 @@ .then((builder) => {

const tooLong = data.concat([0.666]);
model.setInputData(MNIST_IN_NAME, tooLong); // should throw
model.setInputData('bad_input_name', data); // should throw
})
.then(assert.fail, (err) => {
assert.ok(err instanceof Error);
assert.ok(err.message.includes('too long'));
assert.ok(err.message.includes('variable not found'));
});
});
});
describe('#run tests', function () {
it('should throw with invalid arg 1', function () {
it('should throw if input data is too short', function () {
return menoh.create(ONNX_FILE_PATH)

@@ -500,12 +626,12 @@ .then((builder) => {

model.setInputData(MNIST_IN_NAME, data);
model.run('bad')
const tooShort = [0, 1, 2];
model.setInputData(MNIST_IN_NAME, tooShort); // should throw
})
.then(assert.fail, (err) => {
assert.ok(err instanceof Error);
assert.ok(err.message.includes('arg 1'));
assert.ok(err.message.includes('too short'));
});
});
it('second run() on the same model should fail', function () {
it('should throw if input data is too long', function () {
return menoh.create(ONNX_FILE_PATH)

@@ -519,27 +645,13 @@ .then((builder) => {

model.setInputData(MNIST_IN_NAME, data);
let err1 = null;
let err2 = null;
return Promise.all([
model.run().catch((err) => {
err1 = err;
}),
model.run().catch((err) => {
err2 = err;
}),
])
.then(() => {
assert.ok(!err1);
assert.ok(err2 instanceof Error);
assert.ok(err2.message.includes('in progress'));
validateOutput(model, batchSize);
});
const tooLong = data.concat([0.666]);
model.setInputData(MNIST_IN_NAME, tooLong); // should throw
})
.then(assert.fail, (err) => {
assert.ok(err instanceof Error);
assert.ok(err.message.includes('too long'));
});
});
});
describe('#getOutput tests', function () {
describe('#getOutput failure tests', function () {
it('should throw with invalid output name', function () {

@@ -546,0 +658,0 @@ return menoh.create(ONNX_FILE_PATH)

src/model.cpp

Sorry, the diff of this file is not supported yet

src/model.h

Sorry, the diff of this file is not supported yet