tf.cast on native python float to dtype=tf.float64 leads to loss of precision

[st--]

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): no
• OS Platform and Distribution: Linux Ubuntu 18.04
• TensorFlow installed from (source or binary): binary (pip install)
• TensorFlow version (use command below): v2.1.0-rc2-17-ge5bf8de 2.1.0, v2.0.0-rc2-26-g64c3d38 2.0.0
• Python version: 3.7

Describe the current behavior

tf.cast() on a python float (e.g, a literal float constant, not a numpy float/array type) does an implicit conversion to the TensorFlow default dtype tf.float32, which can result in loss of precision when intending to cast to tf.float64:

>>> tf.cast(0.2, tf.float64)
<tf.Tensor: id=37, shape=(), dtype=float64, numpy=0.20000000298023224>

Describe the expected behavior

>>> tf.cast(0.2, tf.float64)
<tf.Tensor: id=37, shape=(), dtype=float64, numpy=0.2>

Code to reproduce the issue
See above.

Other info / logs
This was discovered as a bug in GPflow, which we built a work-around for in GPflow/GPflow#1211 - but this is a pervading issue, and it would be good to fix this upstream instead of having to write and use a gpflow.cast() everywhere just to work around having potentially passed in a python float.

[st--]

The simple workaround would be to add

    if not tf.is_tensor(value):
        return tf.convert_to_tensor(value, dtype)

to the beginning of the definition of tf.cast().

This is related to tensorflow#26033 - not having to support our own "default dtype" implementation would remove the need for a lot of the calls to tf.cast.

[yongtang]

Added PR #35961 for the fix.

[tensorflow-bot]

Are you satisfied with the resolution of your issue?
Yes
No