大規(guī)模深度學(xué)習(xí)，快速處理和解析TFRecord已經(jīng)是必備要求了，記錄一下如何快速預(yù)覽和解析TFRecord
導(dǎo)入相關(guān)包

import tensorflow as tf
from tensorflow.python.ops.parsing_ops import  FixedLenFeature
tf.enable_eager_execution()
tf.logging.set_verbosity(tf.logging.INFO)

加載TFRecord（這里我加載的壓縮格式）

filenames = 'data/20210830/part-r-00000.gz'
raw_dataset = tf.data.TFRecordDataset(filenames=filenames,compression_type='GZIP')

快速解析并預(yù)覽內(nèi)容

for raw_record in raw_dataset.take(1):
    example = tf.train.Example()
    example.ParseFromString(raw_record.numpy())
    print(example)

輸出如下：

features {
  feature {
    key: "album_fea"
    value {
      float_list {
        value: 906.0
        value: 1957.0
      }
    }
  }
  feature {
    key: "albumid"
    value {
      bytes_list {
        value: "41595773"
      }
    }
  }
  feature {
    key: "is_click"
    value {
      float_list {
        value: 0.0
      }
    }
  }
}

上面只是單純的解析為字符串，如果需要解析為可以Tensorflow可以操作的數(shù)據(jù)格式，則需要根據(jù)輸出的數(shù)據(jù)格式定義feature_description

feature_description = {
        'albumid': FixedLenFeature(shape=[1], dtype=tf.string),
        'album_fea': FixedLenFeature(shape=[2], dtype=tf.float32),
        'is_click': FixedLenFeature( shape=[1], dtype=tf.float32)
    }
for serialized_example in raw_dataset.take(1):
    features = tf.io.parse_single_example(serialized_example, feature_description)
    print(features)

輸出如下

{'album_fea': <tf.Tensor: id=748, shape=(2,), dtype=float32, numpy=array([ 906., 1957.], dtype=float32)>,
 'albumid': <tf.Tensor: id=749, shape=(1,), dtype=string, numpy=array([b'41595773'], dtype=object)>,
 'is_click': <tf.Tensor: id=750, shape=(1,), dtype=float32, numpy=array([0.], dtype=float32)>
}

參考地址：
TFRecord 和 tf.Example | TensorFlow Core (google.cn)