tf.keras.metrics.mean_absolute_error exception information is misleading.

[triumph-wangyuyang]

System information.

• Have I written custom code (as opposed to using a stock example script provided in Keras):YES
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):Windows 11
• TensorFlow installed from (source or binary):source
• TensorFlow version (use command below): 2.11
• Python version:3.8
• GPU model and memory:RTX3060
• CUDA/cuDNN version:CUDA: 11.2 cuDNN 8.1

Describe the problem.

The exceptions thrown by tf.keras.metrics.mean_absolute_error(y_true, y_pred) are different when running the following two test codes. In code1, I set y_pred to bool, and the exception message prompts me that bool is not in the legal type (bfloat16, half, float, double, uint8, int8, uint16, int16, int32, int64, complex64, complex128, uint32, uint64) (Remember that it includes uint16, which will be useful later). In code2, I set y_pred to uint16, and the exception information reminded me that uint16 is not in the legal type (bfloat16, half, float, double, int8, int16, int32, int64). It is obvious that the exception information before and after is contradictory. In addition, tf.keras.metrics.mean_absolute_percentage_error, tf.keras.metrics.mape, tf.keras.metrics.MAPE, tf.keras.metrics.mae, tf.keras.metrics.MAE, tf.keras.losses.mean_absolute_error, tf.keras.losses.mean_absolute_percentage_error, tf.keras.losses.mape, tf.keras.losses.MAPE, tf.keras.losses.mae, tf.keras.losses.MAE all have the same doubt.

Standalone code to reproduce the issue.

code 1:

import tensorflow as tf
y_true = tf.random.uniform([2, 3], minval=-5, maxval=5, dtype=tf.int32)
y_pred = tf.cast(tf.random.uniform([2, 3], minval=-5, maxval=5, dtype=tf.int32), dtype=tf.bool)
out = tf.keras.metrics.mean_absolute_error(y_true, y_pred)

code1 result : tensorflow.python.framework.errors_impl.InvalidArgumentError: Value for attr 'T' of bool is not in the list of allowed values: bfloat16, half, float, double, uint8, int8, uint16, int16, int32, int64, complex64, complex128, uint32, uint64; NodeDef: {{node Sub}}; Op<name=Sub; signature=x:T, y:T -> z:T; attr=T:type,allowed=[DT_BFLOAT16, DT_HALF, DT_FLOAT, DT_DOUBLE, DT_UINT8, DT_INT8, DT_UINT16, DT_INT16, DT_INT32, DT_INT64, DT_COMPLEX64, DT_COMPLEX128, DT_UINT32, DT_UINT64]> [Op:Sub]

code 2:

import tensorflow as tf
y_true = tf.random.uniform([2, 3], minval=-5, maxval=5, dtype=tf.int32)
y_pred = tf.cast(tf.random.uniform([2, 3], minval=-5, maxval=5, dtype=tf.int32), dtype=tf.uint16)
out = tf.keras.metrics.mean_absolute_error(y_true, y_pred)

code2 result : tensorflow.python.framework.errors_impl.InvalidArgumentError: Value for attr 'T' of uint16 is not in the list of allowed values: bfloat16, half, float, double, int8, int16, int32, int64; NodeDef: {{node Abs}}; Op<name=Abs; signature=x:T -> y:T; attr=T:type,allowed=[DT_BFLOAT16, DT_HALF, DT_FLOAT, DT_DOUBLE, DT_INT8, DT_INT16, DT_INT32, DT_INT64]> [Op:Abs]

[tilakrayal]

@triumph-wangyuyang, The exception information is auto-generated based on kernel registrations. You will need to update the python docstrings. Thank you!

[triumph-wangyuyang]

update the python docstrings

Thank you for your reply, what is the specific operation, I did not understand the meaning of "update the python docstrings".

[google-ml-butler[bot]]

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

[triumph-wangyuyang]

The exception information is auto-generated based on kernel registrations. You will need to update the python docstrings. Thank you!

Hello, I have changed here, but the problem still exists, can you answer it?

[triumph-wangyuyang]

@triumph-wangyuyang, The exception information is auto-generated based on kernel registrations. You will need to update the python docstrings. Thank you!

Hi, can you explain to me why these problems occur?

[tilakrayal]

@gowthamkpr, I was able to reproduce the issue on tensorflow v2.9, v2.11 and nightly. Kindly find the gist of it here.