Open xypan1232 opened 5 years ago
Replacing softmax cross entropy loss with a sigmoid cross entropy loss should do the job:)
On Sun 23. Sep 2018 at 18:02 Xiaoyong Pan notifications@github.com wrote:
Hi,
I have a multi-label classification problem, where one node can have multi labels, DO I need change the code for multi-class classification? thanks.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tkipf/pygcn/issues/18, or mute the thread https://github.com/notifications/unsubscribe-auth/AHAcYLwttWfGA_6UUaTJlG6Wj73HZtFmks5ud7B9gaJpZM4W1wj7 .
thanks.
I changed the the GCN model return F.sigmoid(x) and replace the loss function F.nll_loss using loss_train = F.binary_cross_entropy(output[idx_train], labels[idx_train]). is my change correct?
Sounds correct!
On Sun 23. Sep 2018 at 18:28 Xiaoyong Pan notifications@github.com wrote:
thanks.
I changed the the GCN model return F.sigmoid(x) and replace the loss function F.nll_loss using loss_train = F.binary_cross_entropy(output[idx_train], labels[idx_train]). is my change correct?
— You are receiving this because you commented.
Reply to this email directly, view it on GitHub https://github.com/tkipf/pygcn/issues/18#issuecomment-423828841, or mute the thread https://github.com/notifications/unsubscribe-auth/AHAcYBdnpghPriJCdJr_vKJNTUcsYUNXks5ud7a9gaJpZM4W1wj7 .
hello, have you finish the multi-label problem?
I need to use GCN with TensorFlow, have you try multi-label problem with Tensorflow? The changes I make as follow: loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=preds, labels=labels) (from metrics.py) correct_prediction = tf.equal(preds, labels). (from metrics.py) return tf.nn.sigmoid(self.outputs). (from models.py)
So is it correct?
@Chengmeng94 It seems to be correct. Make sure to adapt the accuracy function as well to handle a sigmoid output instead of a softmax output. currently it is just taking the max, as you would expect for a softmax function, so it is just evaluating a single label. This is typically done with some threshold.
@anjanaskumar27 My response was to a mult-label situation, rather than a multi-class situation. I think the original code already works for multi-class situations.
@Baukebrenninkmeijer sorry, my bad. I meant to ask procedure for multi label. The following are my changes: loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=preds, labels=labels) and correct_prediction = tf.equal(preds, labels) in metrics.py . And use tf.nn.sigmoid(self.outputs) to predict in models.py. As suggested in https://github.com/tkipf/gcn/issues/119 Can you tell me what is wrong here? Or what needs to be updated in accuracy?
@anjanaskumar27 As far as I can see here, this should be correct except maybe for the metric. I'm not an expert on metrics for multilabel, but I expect just accuracy (which is what you get with the tf.equal) is not the most appropriate, although it still gives you some indication.
@Chengmeng94 @tkipf @Baukebrenninkmeijer , to allow multi-label mulit-class classification:
In addition to @Chengmeng94 changes (as follows), loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=preds, labels=labels) in model.py tf.nn.sigmoid(self.outputs) to predict in models.py
I added, logits_prob = tf.nn.sigmoid(preds, name = 'log_prob') predictions = tf.cast(logits_prob > 0.5, tf.float32,name='predictions') correct_prediction = tf.equal(predictions,labels)
Is this correct?
@anjanaskumar27 tensorflow is not my strongsuit, but a quick google advised the following. Use tf.round()
to push the prediction to either 0 or 1. Then you can calculate metrics. This appears to make sense. Let me know if it works out. Please report if correct_prediction is an NxL matrix, where N is the number of samples and L the number of labels.
So in short, i think you should do:
correct_prediction = tf.equal(tf.round(tf.nn.sigmoid(self.outputs)), labels)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
@Baukebrenninkmeijer , thank you for the suggestion. I am new to tensorflow as well, thus finding it little tricky to modify things. For me, the size of correct predictions is 2x1. I may have to fix this? Additionally, wrt to the code in utils.py, line #77, idx_val = range(len(y), len(y)+500), do you have any idea why they add 500 here? My graph with total nodes of 70, seems to fail because of this. Thank you for the help!
Hi,
I have a multi-label classification problem, where one node can have multi labels, DO I need change the code for multi-class classification? thanks.