如何在 GCP AI Platform 上使用 TFRecord 文件进行批量预测? [英] How to use a TFRecord file for batch prediction on GCP AI Platform?

问题描述

TL;博士Google Cloud AI Platform 在进行批量预测时如何解压 TFRecord 文件?

TL;DR How does Google Cloud AI Platform unpack TFRecord files when doing batch predictions?

我已将经过训练的 Keras 模型部署到 Google Cloud AI Platform，但我在批处理预测的文件格式方面遇到了问题.对于训练，我使用 tf.data.TFRecordDataset 读取 TFRecord 列表，如下所示，一切正常.

I have deployed a trained Keras model to Google Cloud AI Platform, but I'm having trouble with the file format for batch predictions. For training I'm using a tf.data.TFRecordDataset to read a list of TFRecord like the following which all works fine.

def unpack_tfrecord(record):
    parsed = tf.io.parse_example(record, {
        'chunk': tf.io.FixedLenFeature([128, 2, 3], tf.float32),  # Input
        'class': tf.io.FixedLenFeature([2], tf.int64),            # One-hot classification (binary)
    })

    return (parsed['chunk'], parsed['class'])

files = [str(p) for p in training_chunks_path.glob('*.tfrecord')]
dataset = tf.data.TFRecordDataset(files).batch(32).map(unpack_tfrecord)
model.fit(x=dataset, epochs=train_epochs)
tf.saved_model.save(model, model_save_path)

我将保存的模型上传到 Cloud Storage 并在 AI Platform 中创建一个新模型.AI Platform 文档指出Batch with gcloud tool [supports] Text file with JSON instance strings or TFRecord file (may becompressed)"(https://cloud.google.com/ai-platform/prediction/docs/overview#prediction_input_data).但是当我提供 TFRecord 文件时，我收到错误:

I upload the saved model to Cloud Storage and create a new model in AI Platform. AI Platform documentation states that "Batch with gcloud tool [supports] Text file with JSON instance strings or TFRecord file (may be compressed)" (https://cloud.google.com/ai-platform/prediction/docs/overview#prediction_input_data). But when I provide a TFRecord file i get the error:

("'utf-8' codec can't decode byte 0xa4 in position 1: invalid start byte", 8)

我的 TFRecord 文件包含一堆 Protobuf 编码的 tf.train.Example.我没有向 AI Platform 提供 unpack_tfrecord 函数，所以我想它无法正确解包是有道理的，但我知道从这里开始的节点.由于数据太大，我对使用 JSON 格式不感兴趣.

My TFRecord file contains a bunch of Protobuf encoded tf.train.Example. I'm not providing the unpack_tfrecord function to AI Platform, so I guess it makes sense that it can not unpack it properly, but I have node idea where to go from here. I'm not interested in using the JSON format as the data is too large.

推荐答案

我不知道这是否是解决此问题的最佳方式，但对于 TF 2.x，您可以执行以下操作:

I don't know if this is the best way of going about this but for TF 2.x you can do something like:

import tensorflow as tf

def make_serving_input_fn():
    # your feature spec
    feature_spec = {
        'chunk': tf.io.FixedLenFeature([128, 2, 3], tf.float32),  
        'class': tf.io.FixedLenFeature([2], tf.int64),
    }

    serialized_tf_examples = tf.keras.Input(
        shape=[], name='input_example_tensor', dtype=tf.string)

    examples = tf.io.parse_example(serialized_tf_examples, feature_spec)

    # any processing 
    processed_chunks = tf.map_fn(
        <PROCESSING_FN>, 
        examples['chunk'], # ?
        dtype=tf.float32)

    return tf.estimator.export.ServingInputReceiver(
        features={<MODEL_FIRST_LAYER_NAME>: processed_chunks},
        receiver_tensors={"input_example_tensor": serialized_tf_examples}
    )


estimator = tf.keras.estimator.model_to_estimator(
    keras_model=model,
    model_dir=<ESTIMATOR_SAVE_DIR>)

estimator.export_saved_model(
    export_dir_base=<WORKING_DIR>,
    serving_input_receiver_fn=make_serving_input_fn)

这篇关于如何在 GCP AI Platform 上使用 TFRecord 文件进行批量预测?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！