CODEWITHSUNDEEP
pythonmachine-learningtensorflowclassificationmultilabel-classification
guemues
guemues

Reputation: 589

Multi-Label Classifier in Tensorflow

I want to develop a multi-label classifier with TensorFlow, i try to mean there exists multiple label which contains multiple classes. To illustrate you can image the situation like:

I want to classify these two labels with neural network. For now, i used different class label for every (label-1, label-2) pair classes. Which means i have 4 x 4 = 16 different label.

By training my model with

current loss

cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction), reduction_indices=[1])) 

# prediction is sofmaxed
loss = cross_entropy + regul * schema['regul_ratio'] # regul things is for regularization 
train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)

However i think that a multi-label training will work nicer in such a condition.

n_samples of [x1, x2, x3, x4 ...] # features

n_samples of [[0, 0, 0, 1], [0, 0, 1, 0]] # no raining and cloudy

How can i make a softmax probability distribution predictor with tensorflow. Is there any working example for multilabel problem like this one. How will be my loss tensor like?

Upvotes: 4

Views: 3991

Answers (1)

Steven
Steven

Reputation: 5162

Why not just have your network produce two different outputs?

network -> prediction1 and prediction2

where prediction1 and prediction2 are both [#,#,#,#] but what I describe below works even if they're different sizes.

Then just run 

loss1 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(prediction1, labels_1))
loss2 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(prediction2, labels_2))

loss = loss1 + loss2

Upvotes: 6

Related Questions