Tensorflow - Map-Function

Question

I have a question concerning tf's map function. I encounter weird behavior with this function. If I do as stated in manual

label_tensor # shape [150, 1]
y = tf.map_fun(lambda x: x*x, label_tensor)
#  returns y as a tensor with label_tensor x^2

however, if i want to implement my own function it doesn't seem to work. It just always passes a tensor to the specified function which is not made to handle tensors.

y = tf.map_fn(special_fun, label_tensor)

def special_fun(key):
    return int(2000 * round(float(key)/2000))

#  TypeError: float() argument must be a string or a number, not 'Tensor'

I am somehow not seeing the issue here. Also if I try something like: tmp_label_list = tf.Session().run(label_tensor) print(tmp_label_list) # prints out an evaluated list, [1, 2, 3, 3, 1, 2, 2,...] But if I then pass [special_fun(i) for i in tmp_label_list] it raises the Type-Error again, that it expected no 'Tensor'

What am I missing or doing wrong? Thanks in advance.

Answer 1

In tf.map_fn, the given function is expected to accept tensors with the same shape as the given tensor but removing the first dimension (that is, the function will receive each element as a tensor). In any case, what you are trying to do can be done directly (and more efficiently) without using tf.map_fn:

y = tf.cast(2000 * tf.round(tf.cast(key, tf.float32) / 2000), tf.int32)

tf.map_fn is generally reserved for particular cases where vectorization is not possible. However, if you wanted to use it anyway you would have to do something like this:

y = tf.map_fn(special_fun, label_tensor)

def special_fun(key):
    return tf.cast(2000 * tf.round(tf.cast(key, tf.float32) / 2000), tf.int32)

There is also tf.py_func, which lets you apply a regular Python function to a tensor (not to each of its elements in this case, but to the tensor as a whole, given as a NumPy array). This can also be useful for specific cases but you should avoid using it when possible as it is less efficient and also it cannot be serialized.