Sigmoid function returns NaN no matter the input

Question

I am trying to use the tf.math.sigmoid() function on a tensor in my code. I am getting a tensor from the NN (let's call it tensor a) along the lines of:

[[[[-0.8762997 ]
[-0.8903917 ]
[-0.8672142 ]
[-0.8688538 ]
[-0.9621545 ]
[-0.89997154]
[-0.8640675 ]
[-0.93268245]
[-0.8761404 ]
[-0.85574013]
[-0.8742257 ]
[-0.89015967]
[-0.8417985 ]
[-0.8387405 ]
[-0.8407151 ]
[-0.8985772 ]
[-0.8235554 ]
[-0.8261194 ]
[-0.8440901 ]
[-0.9285601 ]
[-0.8315757 ]
[-0.8859896 ]
[-0.94726914]
[-0.8773718 ]
[-0.80427724]
[-0.86090446]
[-0.8038499 ]
[-0.92629945]
[-0.7933995 ]
[-0.7654424 ]]]]

I am then applying a tf.reduce_mean(a, axis=2) yielding:

[[[-0.8665478]]]

of type <class 'tensorflow.python.framework.ops.Tensor'> (let's call this tensor b).

Up to now everything works as expected. However, when I now plug tensor b into the sigmoid function (as such: c = tf.math.sigmoid(b)), tensor c always yields [[[nan]]]!

This happens no matter the values in tensor b. Various values have come out in tensor b (small negative, larger negative, small positive and larger positive).

I've tried various things: squeezing tensor b, recasting it as tf.float32, etc. but the result stays the same! I have also tried casting a constant and plugging it into a sigmoid, which works fine...

Why is this happening? (Thanks so much already for any help!)

** EDIT ** More of the code:

class Model
   def __init__ (self, data)
      # < some code > that computes self.a as a tensor of shape 1x1x30x1
      self.b = tf.reduce_mean(self.a, axis=2)
      self.c = tf.math.sigmoid(self.b)

      x = tf.constant([-128.0], dtype=tf.float32)
      self.testing = tf.math.sigmoid(x) 

if name = '__main__':
   # <load some data>
   model = Model(loaded_data)
   # <compute some cross entropy>
   train_step = tf.compat.v1.train.AdamOptimizer(learning_rate).minimize(cross_entropy, global_step=global_step)
      
   with tf.compat.v1.Session() as sess:
     sess.run(tf.compat.v1.global_variables_initializer())  
     sess.run(tf.compat.v1.local_variables_initializer())

     [a_out, b_out, c_out, test_out] = sess.run([model.a, model.b, model.c, model.testing])

     print(a_out)
     print(b_out)
     print(c_out)
     print(test_out)

The resulting outputs are shown above, except for the test variable, that yields the expected result of:

[0.]