How to implement triplet loss with mxnet

[universewill]

I refer to #1502 and the forum provided can not be opened.
I think a tutorial about triplet loss can be added to the tutorials.

[zihaolucky]

It's simple once you know how to use sym.MakeLoss

    kernels = [(1, feature_size), (2, feature_size), (3, feature_size)]
    for i in range(len(kernels)):
        conv_weight.append(mx.sym.Variable('conv' + str(i) + '_weight'))
        conv_bias.append(mx.sym.Variable('conv' + str(i) + '_bias'))

    fa = get_conv(data=anchor,
                  kernels=kernels, conv_weight=conv_weight, conv_bias=conv_bias,
                  entity_weight=entity_weight, entity_bias=entity_bias,
                  feature_name='fa')  # share weight.
    fs = get_conv(data=same,
                  kernels=kernels, conv_weight=conv_weight, conv_bias=conv_bias,
                  entity_weight=entity_weight, entity_bias=entity_bias,
                  feature_name='fs')
    fd = get_conv(data=diff, 
                  kernels=kernels, conv_weight=conv_weight, conv_bias=conv_bias,
                  entity_weight=entity_weight, entity_bias=entity_bias,
                  feature_name='fd') 

    """
    triple-loss
    """
    fs = fa - fs
    fd = fa - fd
    fs = fs * fs
    fd = fd * fd
    fs = mx.sym.sum(fs, axis=1, keepdims=1)
    fd = mx.sym.sum(fd, axis=1, keepdims=1)
    loss = fd - fs
    loss = one - loss  # a scalar
    loss = mx.sym.Activation(data=loss, act_type='relu')  # acts like a norm.
    triple_loss = mx.sym.MakeLoss(loss)

[zihaolucky]

Refer to https://github.com/xlvector/learning-dl/tree/master/mxnet/triple-loss

[universewill]

@zihaolucky
Thanks for your reply.
I see that the data shape is like this:

    self.provide_data = [('same', (batch_size, 3, 32, 32)), \

                         ('diff', (batch_size, 3, 32, 32)), \
                         ('one', (batch_size, ))]
    self.provide_label = [('anchor', (batch_size, 3, 32, 32))]

I wonder why the anchor is treated as label. Will the model be effected If i change the label to 'one', and put the 'anchor' in data, and change the batch generator accordingly?

[zihaolucky]

The shape is up to you. Check out the example in the link.

[meanmee]

@zihaolucky nice work！

[handong1587]

Hi, @zihaolucky , I think the triple-loss example you mentioned above is very helpful, but I don't think it is a complete tutorial for getting started. Can you kindly guide me how to prepare the training data for that triplet loss? I can see the script is loading some *.npy files but don't know exactly what they are: https://github.com/xlvector/learning-dl/blob/master/mxnet/triple-loss/triplet_loss.py#L155-L160
Thanks very much.

[zjz5250]

@zihaolucky,
Hi, Don't we need to compute the gradient ?

[szha]

@apache/mxnet-committers: This issue has been inactive for the past 90 days. It has no label and needs triage.

For general "how-to" questions, our user forum (and Chinese version) is a good place to get help.

[goswamig]

Please remove the label: Need Triage
Please add the label: Doc

Since the tutorial needs to be updated.