tf.image.cropAndResize 抛出 "method must be bilinear or nearest, but was undefined" 错误

Question

我正在尝试使用 tensorflow.js（特别是 tfjs-node）在 firebase 云函数中运行图像分类模型，但是运行遇到了流动错误：

Error: method must be bilinear or nearest, but was undefined
    at assert (/workspace/node_modules/@tensorflow/tfjs-core/dist/tf-core.node.js:698:15)
    at cropAndResize_ (/workspace/node_modules/@tensorflow/tfjs-core/dist/tf-core.node.js:21340:5)
    at Object.cropAndResize__op [as cropAndResize] (/workspace/node_modules/@tensorflow/tfjs-core/dist/tf-core.node.js:4287:29)
    at prepImage (/workspace/handlers/models.js:58:35)
    at /workspace/handlers/models.js:68:44
    at processTicksAndRejections (internal/process/task_queues.js:97:5)
    at async exports.isFurnished (/workspace/handlers/models.js:10:17)
    at async exports.getanalysis (/workspace/handlers/apis.js:103:16)

tf.image.cropAndResize() 函数正在抛出错误。此错误的奇怪之处在于 cropAndResize() 应自动使用其在 docs.

中指定的默认值“双线性”

奇怪的是，当我在本地运行时，我没有收到任何错误。我的本地机器是运行ning node v12.16.0.

下面是我的代码。请注意，我只是从 firebase 存储中加载 signature.json 并获取/标准化图像（我没有加载和运行实际的 ts 模型）。

const { admin, db } = require("../util/admin");
const firebase = require("firebase");
const tf = require("@tensorflow/tfjs-node");
const fetch = require("node-fetch");

exports.isFurnished = async (imgUrl) => {
  const sigPath = "models/signature.json";

  const signature = await loadSignature(sigPath);
  const image = await loadImage(imgUrl, signature);

  return "It worked!";
};

//signature---------------------------
const loadSignature = (filePath) => {
  let file = admin.storage().bucket().file(filePath);
  return file
    .download()
    .then((res) => JSON.parse(res[0].toString("utf8")))
    .catch((err) => err.message);
};

//Image-------------------------------
const loadImage = (imgUrl, signature) => {
  return fetchImage(imgUrl).then((image) => prepImage(image, signature));
};
const fetchImage = async (url) => {
  const response = await fetch(url);
  const buffer = await response.buffer();
  return buffer;
};
const prepImage = (rawImage, signature) => {
  const image = tf.node.decodeImage(rawImage, 3);

  const [height, width] = signature.inputs.Image.shape.slice(1, 3);
  const [imgHeight, imgWidth] = image.shape.slice(0, 2);
  const normalizedImage = tf.div(image, tf.scalar(255));
  const reshapedImage = normalizedImage.reshape([1, ...normalizedImage.shape]);

  let top = 0;
  let left = 0;
  let bottom = 1;
  let right = 1;
  if (imgHeight != imgWidth) {
    const size = Math.min(imgHeight, imgWidth);
    left = (imgWidth - size) / 2 / imgWidth;
    top = (imgHeight - size) / 2 / imgHeight;
    right = (imgWidth + size) / 2 / imgWidth;
    bottom = (imgHeight + size) / 2 / imgHeight;
  }

  return tf.image.cropAndResize(
    reshapedImage,
    [[top, left, bottom, right]],
    [0],
    [height, width]
  );
};

我是否犯了一个我没有看到的错误，或者这是一个节点 and/or tsjs 问题？

此外，添加“双线性”参数会产生此错误：

Error: Invalid napi_status: A number was expected

Answer 1

我正在使用 bodypix。直到今天早上它都工作正常。虽然我没有做任何更改，但从今天下午开始就出现了这个确切的错误。这可能是 Tensorflow 的问题。或者，

我查看了 Windows 8.1。在那里，它工作得很好。问题发生在 windows 10.

编辑：我很确定它来自 TensorFlow。不是 windows。我正在使用 CDN 获取 bodypix，更新 cdn 地址后错误消失了。

上一个：https://cdn.jsdelivr.net/npm/@tensorflow-models/body-pix/dist/body-pix.min.js https://cdn.jsdelivr.net/npm/@tensorflow/tfjs/dist/tf.min.js

现在：https://cdn.jsdelivr.net/npm/@tensorflow-models/body-pix@2.0.5/dist/body-pix.min.js https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@2.7.0/dist/tf.min.js

Answer 2

正如评论，TensorFlow.js 版本 2.8.0 似乎引入了一些重大变化。解决方法（在撰写本文时）是继续使用 2.7.0 版。

tf.image.cropAndResize 抛出 "method must be bilinear or nearest, but was undefined" 错误

tf.image.cropAndResize throwing "method must be bilinear or nearest, but was undefined" error

node.js

firebase

tensorflow.js