将tfrecords中的原始字节解码为tf.feature_column.numeric_column功能

Question

I have a tfrecords file storing images as bytestrings. I want to define the feature column for this as tf.feature_column.numeric_column("image", shape=[64, 64], dtype=tf.float32) but since it's not stored as a float_list in the tfrecords file this doesn't work.

I then tried to use the normalizer_fn argument of numeric_column which I defined as.

def decode(image_bytestring):
    img = tf.reshape(tf.decode_raw(image_bytestring, tf.uint8), [28, 28])
    img = tf.cast(img, tf.float32)
    return img

...

examples = tf.parse_example(
            serialized_batch,
            tf.feature_column.make_parse_example_spec(feature_columns))

The first problem however is that the parse spec FixedLenFeature(shape=(28, 28), dtype=tf.float32, default_value=None) generated by this feature_column says to parse for a float32 when it's actually stored as a string which causes an error. So the decode function is not used.

Is there a way around this when using tf.feature_column other than storing the image as a float_list in the tfrecord instead?

Seems like having a static type system would have been nice to guarantee the correct type of feature from the mapping function.

Answer 1

Maybe you can just store images as string bytes, and follow the common way to read images?

feature_map = { 'image': tf.FixedLenFeature([], dtype=tf.string,default_value='') }
features = tf.parse_single_example(example_serialized, feature_map)
image_buffer = features['image']
image = tf.image.decode_image(image_buffer, ...)