pooling

[ngragaei]

I want to use gated pooling instead of max pooling, but i do not know how to tune the pooling gate function with the code. any help, please

`def gated_pooling(inputs, filter, size=2, learn_option='l/c'):
"""Gated pooling operation, responsive
Combine max pooling and average pooling in a mixing proportion,
which is obtained from the inner product between the gating mask and the region being
pooled and then fed through a sigmoid:
fgate(x) = sigmoid(wx) fmax(x) + (1-sigmoid(wx)) favg(x)

   arguments:
     inputs: input of shape [batch size, height, width, channels]
     filter: filter size of the input layer, used to initialize gating mask
     size: an integer, width and height of the pooling filter
     learn_option: learning options of gated pooling, include:
                    'l/c': learn a mask per layer/channel
                    'l/r/c': learn a mask per layer/pooling region/channel combined
   return:
     outputs: tensor with the shape of [batch_size, height//size, width//size, channels]

"""
if learn_option == 'l':
    gating_mask = all_channel_connected2d(inputs)
if learn_option == 'l/c':
    w_gated = tf.Variable(tf.truncated_normal([size,size,filter,filter], stddev=2/(size*size*filter*2)**0.5))
    gating_mask = tf.nn.conv2d(inputs, w_gated, strides=[1,size,size,1], padding='VALID')
if learn_option == 'l/r/c':
    gating_mask = locally_connected2d(inputs)

alpha = tf.sigmoid(gating_mask)

x1 = tf.contrib.layers.max_pool2d(inputs=inputs, kernel_size=[size, size], stride=2, padding='VALID')
x2 = tf.contrib.layers.avg_pool2d(inputs=inputs, kernel_size=[size, size],stride=2, padding='VALID')
outputs = tf.add(tf.multiply(x1, alpha), tf.multiply(x2, (1-alpha)))
return outputs`