How does tf.map_fn work?

问题

Look at the demo:

elems = np.array([1, 2, 3, 4, 5, 6])
squares = map_fn(lambda x: x * x, elems)
# squares == [1, 4, 9, 16, 25, 36]

elems = (np.array([1, 2, 3]), np.array([-1, 1, -1]))
alternate = map_fn(lambda x: x[0] * x[1], elems, dtype=tf.int64)
# alternate == [-1, 2, -3]

elems = np.array([1, 2, 3])
alternates = map_fn(lambda x: (x, -x), elems, dtype=(tf.int64, tf.int64))
# alternates[0] == [1, 2, 3]
# alternates[1] == [-1, -2, -3]

I can't understand the second and third.

For the second: I think the result is [2, -1], because the first time x=np.array([1, 2, 3]) and return 1*2, the second time x=np.array([-1, 1, -1]) and return 1*(-1)

For the third: I think the shape of result is (3,2), because the first time x=1 and return (1,-1), the second time x=2 and return (2,-2), the third time x=3 and return (3,-3).

So how does map_fn work?

回答1:

Tensorflow map_fn, from the docs,

map on the list of tensors unpacked from elems on dimension 0.

in this case, the only axis of the input tensor [1,2,3], or [-1,1,-1]. Operations are thus 1*-1,2*1 and 3*-1, and the results are repacked giving you the tensor shape.

回答2:

For the second: I think the result is [2, -1], because the first time x=np.array([1, 2, 3]) and return 1*2, the second time x=np.array([-1, 1, -1]) and return 1*(-1)

In [26]: a = np.array([[1, 2, 3], [2, 4, 1], [5, 1, 7]])
In [27]: b = np.array([[1, -1, -1], [1, 1, 1], [-1, 1, -1]])
In [28]: elems = (a, b)    
In [29]: alternate = map_fn(lambda x: x[0] * x[1], elems, dtype=tf.int64)
In [30]: alternate.eval()
Out[30]: 
array([[ 1, -2, -3],
       [ 2,  4,  1],
       [-5,  1, -7]])

You will see that it is the tensor in 0 dimension of each element in the elems that is applied to the function.

For the third: I think the shape of result is (3,2), because the first time x=1 and return (1,-1), the second time x=2 and return (2,-2), the third time x=3 and return (3,-3).

In [36]: elems = np.array([[1, 2, 3], [4, 5, 1], [1, 6, 1]])
In [37]: alternates = map_fn(lambda x: (x, -x), elems, dtype=(tf.int64, tf.int64))
In [38]: alternates
Out[38]: 
(<tf.Tensor 'map_6/TensorArrayStack/TensorArrayGatherV3:0' shape=(3, 3) dtype=int64>,
 <tf.Tensor 'map_6/TensorArrayStack_1/TensorArrayGatherV3:0' shape=(3, 3) dtype=int64>)
In [39]: alternates[0].eval()
Out[39]: 
array([[1, 2, 3],
       [4, 5, 1],
       [1, 6, 1]])
In [40]: alternates[1].eval()
Out[40]: 
array([[-1, -2, -3],
       [-4, -5, -1],
       [-1, -6, -1]])

To get the results you expected:

In [8]: elems = np.array([[1], [2], [3]])                                                          
In [9]: alternates = map_fn(lambda x: (x, -x), elems, dtype=(tf.int64, tf.int64))
In [10]: sess = tf.InteractiveSession()                                                            
In [11]: alternates[0].eval()
Out[11]: 
array([[1],
       [2],
       [3]])

In [12]: alternates[1].eval()                                                                      
Out[12]: 
array([[-1],
       [-2],
       [-3]])

May this can help you in understanding the map_fn better.

来源：https://stackoverflow.com/questions/45905601/how-does-tf-map-fn-work

标签

tensorflow