如何在tensorflow中的tf.dataset.map中捕获并引发tfrecord错误<；2

使用tfdataset（tensorflow<2.0），我遇到了一个罕见的损坏文件，无法转换为正确的维度。每个tfrecord行都有要读取的图像的文件名和尺寸。我想捕获错误并打印此文件名，以便将其删除

在哪里放置解析器的try-catch以提升文件名

def _parse_fn(example):
    # Define features
    features = {
        'image/filename': tf.io.FixedLenFeature([], tf.string),
        "image/height": tf.FixedLenFeature([], tf.int64),
        "image/width": tf.FixedLenFeature([], tf.int64),

    }

    # Load one example and parse
    example = tf.io.parse_single_example(example, features)

    # Load image from file
    filename = tf.cast(example["image/filename"], tf.string)
    loaded_image = tf.read_file(filename)
    loaded_image = tf.image.decode_image(loaded_image, 3)

    # Reshape to known shape
    image_rows = tf.cast(example['image/height'], tf.int32)
    image_cols = tf.cast(example['image/width'], tf.int32)

    #Wrap in a try catch and report file failure
    try:
        loaded_image = tf.reshape(loaded_image,
                              tf.stack([image_rows, image_cols, 3]),
                              name="cast_loaded_image")
    except tf.errors.InvalidArgumentError as e:
        print("Image filename: {} yielded {}".format(filename, e))

未捕获错误且未打印文件名

File "/apps/tensorflow/1.14.0.cuda10.gpu/lib/python3.7/site-packages/tensorflow/python/client/session.py", line 1458, in __call__
run_metadata_ptr)
tensorflow.python.framework.errors_impl.InvalidArgumentError:
2 root error(s) found.
  (0) Invalid argument: Input to reshape is a tensor with 480000 values, but the requested shape has 259518
     [[{{node cast_loaded_image}}]]
     [[IteratorGetNext]]
     [[replica_3/retinanet/bn2c_branch2a/FusedBatchNorm/ReadVariableOp/_987]]
  (1) Invalid argument: Input to reshape is a tensor with 480000 values, but the requested shape has 259518
     [[{{node cast_loaded_image}}]]
     [[IteratorGetNext]]
0 successful operations.
3 derived errors ignored.

---

一位同事建议：如果没有其他答案，我会接受，我认为这有一定的价值

一个策略是使用

dataset = dataset.apply(tf.data.experimental.ignore_errors())

并应用返回每个tfrecord的文件名的解析。运行此命令并将其和原始记录的长度进行比较后，您可以找到哪些图像已损坏。我认为肯定还有其他解决方案，但如果您可以访问打包到tfrecords中的原始图像集，则可以区分列表并获取缺少的文件名

dataset = dataset.apply(tf.data.experimental.ignore_errors())