Check if training from scratch or retraining an existing model is needed.

train-function.js

// TFjs and gcs are using a lot of memory, so increasing it will speed up the functions
exports.train = functions
  .runWith({ memory: "2GB" })
  .https.onRequest(async (request, response) => {
    const existJSON = await bucket.file("model.json").exists().then(ex => ex[0]);
    const existBIN = await bucket.file("weights.bin").exists().then(ex => ex[0]);

    if (!existJSON || !existBIN) {
      const model = tf.sequential();
      model.add(tf.layers.dense({ units: 2, inputShape: [vocabulary.length] }));
      model.add(tf.layers.dense({ units: 1, inputShape: [2] }));
      model.compile({ loss: "meanSquaredError", optimizer: "sgd" });

      await db.collection("comments").get()
        .then(async querySnapshot => {
          // Get tensor-like arrays from Firestore
          return await trainSave(model, querySnapshot);
        });
    } else {
      // Download the files if they exist
      await bucket.file("model.json").download({ destination: join(tmpdir, "model.json") });
      await bucket.file("weights.bin").download({ destination: join(tmpdir, "weights.bin") });

      const model = await tf.loadModel("file://" + join(tmpdir, "model.json"));
      model.compile({ loss: "meanSquaredError", optimizer: "sgd" });

      const lastUpdated = await bucket.file("weights.bin").getMetadata()
        .then(metadata => new Date(metadata[0].updated));

      await db.collection("comments").where("publishedAt", ">", lastUpdated).get()
        .then(async querySnapshot => {
          return await trainSave(model, querySnapshot);
        });
    }

    response.send("Success");
  });