keras: Many metrics cannot handle predictions out of [0..1] range

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 10 Pro
• TensorFlow installed from (source or binary): Binary
• TensorFlow version (use command below): v2.7.0-rc1-69-gc256c071bb2 2.7.0
• Python version: 3.9.6
• Bazel version (if compiling from source): N/A
• GPU model and memory: N/A
• Exact command to reproduce: N/A

Describe the problem.

When using BinaryCrossentropy(from_logits=True) loss, which is the non-default but recommended setting per BinaryCrossentropy documentation, looks like prediction values can be out of [0…1] range and this breaks many metrics.

Below list of metrics are known to be broken:

keras.metrics.TruePositives
keras.metrics.FalsePositives
keras.metrics.TrueNegatives
keras.metrics.FalseNegatives
keras.metrics.Precision
keras.metrics.Recall
keras.metrics.AUC

With any of these metrics in place, model.fit(..) call fails with the below error:

tensorflow.python.framework.errors_impl.InvalidArgumentError:  assertion failed: [predictions must be <= 1]  

or

tensorflow.python.framework.errors_impl.InvalidArgumentError:  assertion failed: [predictions must be >= 0]  

Traceback is in the logs section.

Describe the current behavior.

model.fit(..) call fails with BinaryCrossentropy(from_logits=True) loss and any of the metrics listed above.

Describe the expected behavior.

model.fit(..) call succeds.

Standalone code to reproduce the issue.

Below code demonstrates the issue:

import numpy as np
import tensorflow as tf
from tensorflow import keras

x_train = np.random.rand(50,5)
y_train = np.random.rand(50,)

model = tf.keras.models.Sequential([
  tf.keras.layers.Dense(50, activation='relu', name='hidden_1', input_shape=(x_train.shape[1],)),
  tf.keras.layers.Dense(1, activation=None, name='output')
])

METRICS = [
      keras.metrics.Precision(name='precision'),
]

model.compile(optimizer=tf.keras.optimizers.Adam(),
  loss=tf.keras.losses.BinaryCrossentropy(from_logits=True), 
  metrics=METRICS)

model.fit(x_train, y_train, epochs=20)

** Additional Info**

Please see below post in the TensorFlow forum for a discussion: https://discuss.tensorflow.org/t/metrics-related-predictions-must-be-1-error/6144

Below closed issue looks relevant: https://github.com/tensorflow/tensorflow/issues/42182

Source code / logs.

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\utils\traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\tensorflow\python\eager\execute.py", line 58, in quick_execute
    tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
tensorflow.python.framework.errors_impl.InvalidArgumentError:  assertion failed: [predictions must be <= 1] [Condition x <= y did not hold element-wise:] [x (sequential_4/output/BiasAdd:0) = ] [[1.00585222][1.00123906][0.880351603]...] [y (Cast_3/x:0) = ] [1]
         [[node assert_less_equal/Assert/AssertGuard/Assert
 (defined at C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\utils\metrics_utils.py:612)
]] [Op:__inference_train_function_29922]

Errors may have originated from an input operation.
Input Source operations connected to node assert_less_equal/Assert/AssertGuard/Assert:
In[0] assert_less_equal/Assert/AssertGuard/Assert/assert_less_equal/All:
In[1] assert_less_equal/Assert/AssertGuard/Assert/data_0:
In[2] assert_less_equal/Assert/AssertGuard/Assert/data_1:
In[3] assert_less_equal/Assert/AssertGuard/Assert/data_2:
In[4] assert_less_equal/Assert/AssertGuard/Assert/sequential_4/output/BiasAdd:
In[5] assert_less_equal/Assert/AssertGuard/Assert/data_4:
In[6] assert_less_equal/Assert/AssertGuard/Assert/Cast_3/x:

Operation defined at: (most recent call last)
>>>   File "<stdin>", line 1, in <module>
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\utils\traceback_utils.py", line 64, in error_handler
>>>     return fn(*args, **kwargs)
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\engine\training.py", line 1216, in fit
>>>     tmp_logs = self.train_function(iterator)
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\engine\training.py", line 878, in train_function
>>>     return step_function(self, iterator)
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\engine\training.py", line 867, in step_function
>>>     outputs = model.distribute_strategy.run(run_step, args=(data,))
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\engine\training.py", line 860, in run_step
>>>     outputs = model.train_step(data)
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\engine\training.py", line 817, in train_step
>>>     self.compiled_metrics.update_state(y, y_pred, sample_weight)
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\engine\compile_utils.py", line 460, in update_state
>>>     metric_obj.update_state(y_t, y_p, sample_weight=mask)
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\utils\metrics_utils.py", line 73, in decorated
>>>     update_op = update_state_fn(*args, **kwargs)
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\metrics.py", line 177, in update_state_fn
>>>     return ag_update_state(*args, **kwargs)
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\metrics.py", line 1069, in update_state
>>>     return metrics_utils.update_confusion_matrix_variables(
>>>
>>>   File "C:\Users\hakan\AppData\Roaming\Python\Python39\site-packages\keras\utils\metrics_utils.py", line 612, in update_confusion_matrix_variables
>>>     tf.compat.v1.assert_less_equal(
>>>

Function call stack:
train_function -> assert_less_equal_Assert_AssertGuard_false_28872