index.js
104 lines
const functions = require('firebase-functions');
const express = require('express');
const express = require('express');
const Busboy = require('busboy');
const Busboy = require('busboy');
const path = require('path');
const path = require('path');
const tf = require('@tensorflow/tfjs-node');
const tf = require('@tensorflow/tfjs-node');
const labels = require('./model/new_object_detection_1/assets/labels.json');
const labels = require('./model/new_object_detection_1/assets/labels.json');
const app = express();
const app = express();
let objectDetectionModel;
let objectDetectionModel;
async function loadModel() {
async function loadModel() {
// Warm up the model
// Warm up the model
if (!objectDetectionModel) {
if (!objectDetectionModel) {
// Load the TensorFlow SavedModel through tfjs-node API. You can find more
// Load the TensorFlow SavedModel through tfjs-node API. You can find more
// details in the API documentation:
// details in the API documentation:
// https://js.tensorflow.org/api_node/1.3.1/#node.loadSavedModel
// https://js.tensorflow.org/api_node/1.3.1/#node.loadSavedModel
objectDetectionModel = await tf.node.loadSavedModel(
objectDetectionModel = await tf.node.loadSavedModel(
'./model/new_object_detection_1', ['serve'], 'serving_default');
'./model/new_object_detection_1', ['serve'], 'serving_default');
}
}
const tempTensor = tf.zeros([1, 2, 2, 3]).toInt();
const tempTensor = tf.zeros([1, 2, 2, 3]).toInt();
objectDetectionModel.predict(tempTensor);
objectDetectionModel.predict(tempTensor);
}
}
app.get('/', async (req, res) => {
app.get('/', async (req, res) => {
res.sendFile(path.join(__dirname + '/index.html'));
res.sendFile(path.join(__dirname + '/index.html'));
loadModel();
})
})
app.post('/predict', async (req, res) => {
app.post('/predict', async (req, res) => {
// Receive and parse the image from client side, then feed it into the model
// Receive and parse the image from client side, then feed it into the model
// for inference.
// for inference.
const busboy = new Busboy({headers: req.headers});
const busboy = new Busboy({headers: req.headers});
let fileBuffer = new Buffer('');
let fileBuffer = Buffer.from('')
req.files = {file: []};
const files = [];
busboy.on('field', (fieldname, value) => {
req.body[fieldname] = value;
});
busboy.on('file', (fieldname, file, filename, encoding, mimetype) => {
busboy.on('file', (fieldname, file, filename, encoding, mimetype) => {
file.on('data', (data) => {fileBuffer = Buffer.concat([fileBuffer, data])});
file.on('data', (data) => {
fileBuffer = Buffer.concat([fileBuffer, data]);
});
file.on('end', () => {
file.on('end', () => {
const file_object = {
const file_object = {
fieldname,
fieldname,
'originalname': filename,
'originalname': filename,
encoding,
encoding,
mimetype,
mimetype,
buffer: fileBuffer
buffer: fileBuffer
};
};
req.files.file.push(file_object)
files.push(file_object)
});
});
});
});
busboy.on('finish', async () => {
busboy.on('finish', async () => {
const buf = req.files.file[0].buffer;
const buf = files[0].buffer;
const uint8array = new Uint8Array(buf);
const uint8array = new Uint8Array(buf);
loadModel();
// Decode the image into a tensor.
// Decode the image into a tensor.
const imageTensor = await tf.node.decodeImage(uint8array);
const imageTensor = await tf.node.decodeImage(uint8array);
const input = imageTensor.expandDims(0);
const input = imageTensor.expandDims(0);
// Feed the image tensor into the model for inference.
// Feed the image tensor into the model for inference.
const startTime = tf.util.now();
const startTime = tf.util.now();
let outputTensor = objectDetectionModel.predict({'x': input});
let outputTensor = objectDetectionModel.predict({'x': input});
// Parse the model output to get meaningful result(get detection class and
// Parse the model output to get meaningful result(get detection class and
// object location).
// object location).
const scores = await outputTensor['detection_scores'].arraySync();
const scores = await outputTensor['detection_scores'].arraySync();
const boxes = await outputTensor['detection_boxes'].arraySync();
const boxes = await outputTensor['detection_boxes'].arraySync();
const names = await outputTensor['detection_classes'].arraySync();
const names = await outputTensor['detection_classes'].arraySync();
const endTime = tf.util.now();
const endTime = tf.util.now();
outputTensor['detection_scores'].dispose();
outputTensor['detection_scores'].dispose();
outputTensor['detection_boxes'].dispose();
outputTensor['detection_boxes'].dispose();
outputTensor['detection_classes'].dispose();
outputTensor['detection_classes'].dispose();
outputTensor['num_detections'].dispose();
outputTensor['num_detections'].dispose();
const detectedBoxes = [];
const detectedBoxes = [];
const detectedNames = [];
const detectedNames = [];
for (let i = 0; i < scores[0].length; i++) {
for (let i = 0; i < scores[0].length; i++) {
if (scores[0][i] > 0.3) {
if (scores[0][i] > 0.3) {
detectedBoxes.push(boxes[0][i]);
detectedBoxes.push(boxes[0][i]);
detectedNames.push(labels[names[0][i]]);
detectedNames.push(labels[names[0][i]]);
}
}
}
}
res.send({
res.send({
boxes: detectedBoxes,
boxes: detectedBoxes,
names: detectedNames,
names: detectedNames,
inferenceTime: endTime - startTime
inferenceTime: endTime - startTime
});
});
});
});
busboy.end(req.rawBody);
req.pipe(busboy);
req.pipe(busboy);
});
});
loadModel();
loadModel().then(function(){
console.log(`loaded saved model`);
exports.app = functions.https.onRequest(app);
const port = process.env.PORT || 8080;
app.listen(port, () => {
console.log(`listening on port ${port}`);
});
});