Tensorflow 2：序列化和解码时形状不匹配

Question

I have a tensor A of shape (300,256,256). I want to serialize A to save as tfrecord format. But I cannot convert it back to tensor with same shape.

A = tf.convert_to_tensor( *a numpy array with float32 type* )
B = tf.io.serialize_tensor(A)
C = tf.reshape(tf.io.decode_raw(B, out_type=tf.float32),[300,256,256])

If I run the code above, I got a shape error:

tensorflow.python.framework.errors_impl.InvalidArgumentError: Input to reshape is a tensor with 19660806 values, but the requested shape has 19660800 [Op:Reshape]

It seems that when I serialize or when I decode, 6 floats are added. (very weird)

Answer 1

Try using: tf.io.parse_tensor() , instead of tf.io.decode_raw() .

https://www.tensorflow.org/api_docs/python/tf/io/parse_tensor