TFJS Test for EfficientDet Models

test.js

const fs = require('fs');
const path = require('path');
const log = require('@vladmandic/pilogger');
const tf = require('@tensorflow/tfjs');
const tfnode = require('@tensorflow/tfjs-node');

function getTensorFromImage(path) {
  if (!fs.existsSync(path)) return null;
  const data = fs.readFileSync(path);
  const tfimage = tfnode.node.decodeImage(data);
  const expanded = tf.expandDims(tfimage, 0);
  log.info('Image:', path, tfimage.size, 'bytes with shape:', expanded.shape);
  tf.dispose(tfimage);
  return expanded;
}

async function processSavedModel(image, modelPath) {
  log.info('Loading saved model:', modelPath);
  log.data('Model signature:', await tfnode.node.getMetaGraphsFromSavedModel(modelPath))
  const model = await tfnode.node.loadSavedModel(modelPath);
  log.state('TensorFlow/JS Memory', tf.memory());
  try {
    const res = await model.predict(image, {score: 0.1, iou: 0.1, topk: 20});
    log.data('Detected', res);
  } catch (err) {
    log.error(err.stack ? err.stack : err);
  }
}

async function processGraphModel(image, modelPath) {
  log.info('Loadeding layers model:', modelPath);
  const model = await tf.loadGraphModel(`file://${path.join(__dirname, modelPath)}`);
  log.info('Loaded model:', modelPath, model.artifacts.format, model.artifacts.generatedBy, model.artifacts.convertedBy);
  log.data('Model signature:', model.executor._signature);
  log.state('TensorFlow/JS Memory', tf.memory());
  try {
    const res = await model.executeAsync(image);
    log.data('Detected', res);
  } catch (err) {
    log.error(err.stack ? err.stack : err);
  }
}

async function processLayersModel(image, modelPath) {
  log.info('Loadeding layers model:', modelPath);
  const model = await tf.loadLayersModel(`file://${path.join(__dirname, modelPath)}`);
  log.data('Model signature:', model.executor._signature);
  log.state('TensorFlow/JS Memory', tf.memory());
  try {
    const res = await model.executeAsync(image);
    log.data('Detected', res);
  } catch (err) {
    log.error(err.stack ? err.stack : err);
  }
}

async function main() {
  log.header();
  log.info('TensorFlow/JS Version', tf.version);
  await tf.setBackend('tensorflow'); // 'tensorflow' is tfjs-node backend
  await tf.enableProdMode();
  tf.ENV.set('DEBUG', false);
  tf.ENV.set('WEBGL_FORCE_F16_TEXTURES', true);
  await tf.ready();
  log.info('TensorFlow/JS Backend', tf.getBackend());
  log.info('TensorFlow/JS Flags', tf.ENV.flags);

  const image = getTensorFromImage('test.jpg');

  // 1 - saved model from tfhub
  // download: https://tfhub.dev/tensorflow/efficientdet/d0/1?tf-hub-format=compressed
  // error: Session fail to run with error: Expects arg[0] to be uint8 but int32 is provided
  // await processSavedModel(image, 'd0/tfhub'); 

  // 2 - saved model from tf model zoo (should be the same as (1), but file size does not match)
  // download: http://download.tensorflow.org/models/object_detection/tf2/20200711/efficientdet_d0_coco17_tpu-32.tar.gz
  // error: Session fail to run with error: Expects arg[0] to be uint8 but int32 is provided
  // await processSavedModel(image, 'd0/zoo'); 

  // 3 - saved model converted from checkpoint from google's repository (there is no saved model)
  // download: https://storage.googleapis.com/cloud-tpu-checkpoints/efficientdet/coco2/efficientdet-d0.tar.gz
  // convert ckpt to saved: model_inspect.py --runmode=saved_model --model_name=efficientdet-d0 --ckpt_path=ckpt --saved_model_dir saved
  // error: Create kernel failed: Invalid argument: NodeDef mentions attr 'allowed_devices' not in Op<name=VarHandleOp
  // await processSavedModel(image, 'd0/google'); 

  // 4 - graph model converted from (1): tensorflowjs_converter --input_format=tf_saved_model --output_format=tfjs_graph_model . graph
  // error: Error in matMul: inputs must have the same rank of at least 2, got ranks 5 and 2.
  // await processGraphModel(image, 'd0/tfhub/graph/model.json');

  // 5 - graph model converted from (2): tensorflowjs_converter --input_format=tf_saved_model --output_format=tfjs_graph_model . graph
  // error: Error in matMul: inputs must have the same rank of at least 2, got ranks 5 and 2.
  // await processGraphModel(image, 'd0/zoo/graph/model.json');

  // 6 - graph model converted from (3): tensorflowjs_converter --input_format=tf_saved_model --output_format=tfjs_graph_model . graph
  // error: the input tensors shape does not match
  // await processGraphModel(image, 'd0/google/graph/model.json');

  // 7 - graph model converted from (3): tensorflowjs_converter --input_format tf_frozen_model --output_format tfjs_graph_model --output_node_names=detections efficientdet-d0_frozen.pb frozen
  // error: The shape of dict['image_files'] provided in model.execute(dict) must be [-1], but was [1,3268,5843,3]
  // await processGraphModel(image, 'd0/google/frozen/model.json');

  // 8 - layers model converted from keras model from google (seems its weights only, not a full model)
  // download: https://storage.googleapis.com/cloud-tpu-checkpoints/efficientdet/coco2/efficientdet-d0.h5
  // convert: tensorflowjs_converter --input_format keras ./efficientdet-d0.h5 out
  // error: TypeError: Cannot read property 'model_config' of null
  // await processLayersModel(image, 'd0/keras/layers/model.json');

  tf.dispose(image);
}

main();