对数据集的shuffle处理需要设置相应的buffer_size参数，相当于需要将相应数目的样本读入内存，且这部分内存会在训练过程中一直保持占用。完全的shuffle需要将整个数据集读入内存，这在大规模数据集的情况下是不现实的，故需要结合设备内存以及Batch大小将TFRecord文件随机划分为多个子文件，再对数据集做local shuffle（即设置相对较小的buffer_size，不小于单个子文件的样本数）。
Shuffle和划分
下文以一个异常检测数据集（正负样本不平衡）为例，在生成第一批TFRecord时，我将正负样本分别写入单独的TFrecord文件以备后续在对正负样本有不同处理策略的情况下无需再解析example_proto。比如在以下代码中，我对正负样本有不同的验证集比例，并将他们写入不同的验证集文件。
import numpy as np import tensorflow as tf from tqdm.notebook import tqdm as tqdm # TFRecord划分 raw_normal_dataset = tf.data.TFRecordDataset("normal_16_256.tfrecords","GZIP") raw_anomaly_dataset = tf.data.TFRecordDataset("anomaly_16_256.tfrecords","GZIP") normal_val_writer = tf.io.TFRecordWriter(r'ex_1/'+'normal_val_16_256.tfrecords',"GZIP") anomaly_val_writer = tf.io.TFRecordWriter(r'ex_1/'+'anomaly_val_16_256.tfrecords',"GZIP") train_writer_list = [tf.io.TFRecordWriter(r'ex_1/'+'train_16_256_{}.tfrecords'.format(i),"GZIP") for i in range(SUBFILE_NUM+1)] with tqdm(total=LEN_NORMAL_DATASET+LEN_ANOMALY_DATASET) as pbar: for example_proto in raw_normal_dataset: # 划分训练集和测试集 if np.random.random() > 0.99: # 正样本测试集的比例 normal_val_writer.write(example_proto.numpy()) else: train_writer_list[np.random.randint(0,SUBFILE_NUM+1)].write(example_proto.numpy()) pbar.update(1) for example_proto in raw_anomaly_dataset: # 划分训练集和测试集 if np.random.random() > 0.7: # 负样本测试集的比例 anomaly_val_writer.write(example_proto.numpy()) else: train_writer_list[np.random.randint(0,SUBFILE_NUM+1)].write(example_proto.numpy()) pbar.update(1) normal_val_writer.close() anomaly_val_writer.close() for train_writer in train_writer_list: train_writer.close() 读取
raw_train_dataset = tf.data.TFRecordDataset([r'ex_1/'+'train_16_256_{}.tfrecords'.format(i) for i in range(SUBFILE_NUM+1)],"GZIP") raw_train_dataset = raw_train_dataset.shuffle(buffer_size=100000).batch(BATCH_SIZE) parsed_train_dataset = raw_train_dataset.map(map_func=map_func) raw_normal_val_dataset = tf.data.TFRecordDataset(r'ex_1/'+'normal_val_16_256.tfrecords',"GZIP") raw_anomaly_val_dataset = tf.data.TFRecordDataset(r'ex_1/'+'anomaly_val_16_256.tfrecords',"GZIP") parsed_nomarl_val_dataset = raw_normal_val_dataset.batch(BATCH_SIZE).map(map_func=map_func) parsed_anomaly_val_dateset = raw_anomaly_val_dataset.batch(BATCH_SIZE).map(map_func=map_func) 
