如何对 tensorflow 中的 CNN kernel 进行限制？

Question

Our specific training problem requires that all the weights in a CNN kernel must be positive. Anyone knows how to do that?

Answer 1

You are looking for kernel constraints, this link here has the answer to your question. You can either use kernel constraints or simply adopt the lagrange optimization technique to add a term to your overall loss function that removes the feature of the weights that you don't need. Something like this (as it is mentioned in the referenced answer),

added_loss = -tf.minimum(tf.reduce_min(theta),0)
final_loss = your_loss + added_loss
optimizer.minimize(final_loss)

It is also necessary to note that, adding this loss only encourages the optimization to look for a solution which has theta or as you say weights to be all positive, there is no guarantee that the best fit parameters are all positive. To ensure such a behaviour, you can clip the weights between 0 and +infinity.