conversion to tensorflow.js

[jan-bogaerts]

Hi,
I've been trying to convert the mtcnn model to tensorflow.js, but so far, I keep getting empty results back. Do you have any idea what I'm doing wrong?

here's my test code snippet:

const tf = require('@tensorflow/tfjs-node');
const fs = require('fs');

const MIN_SIZE = 40;
const FACTOR = 0.7;
const THRESHOLDS = [0.6, 0.7, 0.8];

async function load() {
    await tf.ready();
    const MODEL_URL = tf.io.fileSystem('./src/models/mtcnn_web/model.json');
    const model = await tf.loadGraphModel(MODEL_URL);
    return model;
}

load().then( async (model) => {
    const image = fs.readFileSync('./test_data/test_image2.jpg');
    const tensor = tf.node.decodeImage(image);
    const floatTensor = tf.cast(tensor, 'float32');
    const input = {
        'input': floatTensor,
        'min_size': tf.tensor(MIN_SIZE),
        'thresholds': tf.tensor(THRESHOLDS),
        'factor': tf.tensor(FACTOR),
    }
    const result = await model.executeAsync(input);
    const values = result[0].dataSync();
    const arr = Array.from(values);
    console.log(arr);
}).catch((error) => {
    console.log(error.toString());
})

• I first converted the model to RGB mode as you mentioned here: Detection differences to facenet-mtcnn #5
• next, I converted the model using this command (details here: https://github.com/tensorflow/tfjs/tree/master/tfjs-converter):
  tensorflowjs_converter --input_format=tf_frozen_model --output_node_names='prob,landmarks,box' /python/mtcnn.pb /python/mtcnn_web
  also tried:
  tensorflowjs_converter --input_format=tf_frozen_model --output_node_names='prob:0,landmarks:0,box:0' /python/mtcnn.pb /python/mtcnn_web
• I also tried with inputs like 'input:0', 'min_size:0'
  I'm not getting any errors, but nothing is detected either. So I'm thinking, maybe somehting is wrong with the model conversion?
  One thing I'm thinking about: I currently presume that the input format is 'frozen_model' cause of the .pb extension?

Thanks.