[TF 2.0] Using tf.py_function which has tf.string type input in dataset.map() generates dtype warning.

[KichangKim]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow):
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Windows 10 x64 1809
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device:
• TensorFlow installed from (source or binary): pip
• TensorFlow version (use command below): unknown 2.0.0-dev20190428
• Python version: 3.6.7
• Bazel version (if compiling from source):
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version: CUDA 10.0, cudnn-10.0-windows10-x64-v7.5.0.56
• GPU model and memory: GeForce GTX 1070 8GB

Describe the current behavior
Using tf.py_function which has tf.string type input generates warning like this:

W0429 14:24:18.965364 13252 backprop.py:820] The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.string

This warning did not shown with v1.12.0-9492-g2c319fb415 2.0.0-alpha0, but 2.0.0-dev20190428 does.

Describe the expected behavior
dtype warning should not be shown.

Code to reproduce the issue

import tensorflow as tf


def transform_tag_python(x):
    return 1.0


dataset = tf.data.Dataset.from_tensor_slices(['tag1'])
dataset = dataset.map(lambda x: tf.py_function(
    transform_tag_python, (x,), (tf.float32,)))

for sample in dataset:
    print(sample)

[shashvatshahi1998]

Look where tf.py_function is defined there is no such warning file present regarding dtype to be float. So why it is showing this?

[rbrueckner]

I am facing the same issue.

[jvishnuvardhan]

@KichangKim There was no warning when TF2.0.0-alpha0 or tf-nightly was used. The issue was reproduced with tf-nightly-2.0-preview. We will check for the source of error and resolve it. Thanks for finding this. Thanks!

[sleighsoft]

Here:

tensorflow/tensorflow/python/eager/backprop.py

Line 841 in 88ed977

logging.WARN, "The dtype of the watched tensor must be "

Having the same issue even though the statement is not within tf.GradientTape.

Even simpler example:

import tensorflow as tf

tensor = tf.constant([1,2,3])
tf.py_function(lambda x: x.numpy(), [tensor], tensor.dtype)

I am wondering why this code is called anyways. My example does not even use GradientTape.

Callstack of test.py (just includes code above):

I am not sure it is intended that a py_function has a GradientTape by default.

[alextp]

It's likely the tape.watch here

tensorflow/tensorflow/python/ops/script_ops.py

Line 108 in 88520c4

tape.watch(tensor)

which needs to be called only for floating dtypes.

[alextp]

I'll happily accept a PR to fix this.

[feihugis]

The PR #30107 is submitted to fix this issue.

cc @alextp

[vineetred]

Looks like the PR failed a few checks. Is the issue still persisting?

[shubhraaich]

This issue still exists in tf-2 version beta-1.

[sleighsoft]

You have to use nightly builds