mean field update

[abeyang00]

During mean field update part,

You seem to concatenate x_s and x_S
Instead of concatenation, aren't you supposed to

before normalize with sigmoid function?

why are you concatenating??

[dontLoveBugs]

Yes, but the implementation of mean field upate in official code is to concatenate x_s and x_S, which is different from Sec 3.2.3 in the paper.
The implementation in official code using caffe framework is as follows:

def MeanFieldUpdate(n, bottom_send, bottom_receive, feat_ind, mf_iter, feat_num):
    '''
    Meanfield updating for the features and the attention for one pair of features.
    bottom_list is a list of observation features derived from the backbone CNN.
    '''
    #generating an attention map
    concat_f = 'concat_f{}_mf{}'.format(feat_ind, mf_iter)
    conv_f = 'conv_f{}_mf{}'.format(feat_ind, mf_iter)
    atten_f = 'atten_f{}_mf{}'.format(feat_ind, mf_iter)
    norm_atten_f = 'norm_atten_f{}_mf{}'.format(feat_ind, mf_iter)
    message_f = 'message_f{}_mf{}'.format(feat_ind, mf_iter)
    filter_message_f = 'filter_message_f{}_mf{}'.format(feat_ind, mf_iter)
    message_scaled = 'message_scaled_f{}_mf{}'.format(feat_ind, mf_iter)
    updated_f = 'updated_f{}_mf{}'.format(feat_ind, mf_iter)

    n[concat_f] = L.Concat(bottom_send, bottom_receive)
    #specify parameter names to make them share between different meanfield updating
    n[atten_f] = L.Convolution(n[concat_f], num_output=feat_num, kernel_size=3, stride=1, pad=1, param=[dict(name='atten_f{}_w'.format(feat_ind), lr_mult=1, decay_mult=1), dict(name='atten_f{}_b'.format(feat_ind), lr_mult=2, decay_mult=0)])
    n[norm_atten_f] = L.Sigmoid(n[atten_f])
    n[message_f] = L.Convolution(bottom_send, num_output=feat_num, kernel_size=3, stride=1, pad=1, param=[dict(name='message_f{}_w'.format(feat_ind), lr_mult=1, decay_mult=1), dict(name='message_f{}_b'.format(feat_ind), lr_mult=2, decay_mult=0)])
    n[filter_message_f] = L.Eltwise(n[message_f], n[norm_atten_f], operation=P.Eltwise.PROD)
    #scale the messages before adding 
    n[message_scaled] = L.Scale(n[filter_message_f], bias_term=True, in_place=True)
    n[updated_f] = L.Eltwise(bottom_receive, n[message_scaled], operation=P.Eltwise.SUM)  

[abeyang00]

Thank you,

also additonal question i had was
you seem to update Mean field 5 times in the code,

Will that be necessary for training too?
was number 5 picked randomly?

[dontLoveBugs]

I think the iteration number is obtained experimentally and it should be a balance between performance and computational efficiency.