How to slice a batch and apply an operation on each slice in TensorFlow

问题

I am a beginner with TensorFlow, I am trying to implement a function that takes a batch as input. It has to slice this batch into several ones, apply some operations on them, then concatenate them to build a new tensor to return. Through my readings, I found there are some implemented function like input_slice_producer and batch_join but I didn't get to work with them. I attached what I've found as solution below, but it's kinda slow, not proper and incapable of detecting the current size of batch. Does any know a better way of doing this?

def model(x):

    W_1 = tf.Variable(tf.random_normal([6,1]),name="W_1")
    x_size = x.get_shape().as_list()[0]
    # x is a batch of bigger input of shape [None,6], so I couldn't 
    # get the proper size of the batch when feeding it 
    if x_size == None:
        x_size= batch_size
    #intialize the y_res
    dummy_x = tf.slice(x,[0,0],[1,6])
    result = tf.reduce_sum(tf.mul(dummy_x,W_1))
    y_res = tf.zeros([1], tf.float32)
    y_res = result
    #go throw all slices and concatenate them to get result
    for i in range(1,x_size): 
        dummy_x = tf.slice(x,[i,0],[1,6])
        result = tf.reduce_sum(tf.mul(dummy_x,W_1))
        y_res = tf.concat(0, [y_res, result])

    return y_res

回答1:

The TensorFlow function tf.map_fn(fn, elems) allows you to apply a function (fn) to each slice of a tensor (elems). For example, you could express your program as follows:

def model(x):
    W_1 = tf.Variable(tf.random_normal([6, 1]), name="W_1")

    def fn(x_slice):
        return tf.reduce_sum(x_slice, W_1)

    return tf.map_fn(fn, x)

It may also be possible to implement your operation more concisely using broadcasting on the tf.mul() operator, which uses NumPy broadcasting semantics, and the axis argument to tf.reduce_sum().

来源：https://stackoverflow.com/questions/35575982/how-to-slice-a-batch-and-apply-an-operation-on-each-slice-in-tensorflow