Felix Lau

Also related is to allow the Python bindings to programmatically delete data. I think it would be helpful to first switch to using something like http://typicode.github.io/lowdb/ or http://brian.io/lawnchair/ for the...

Would you consider an API to specify the weights per label instead of per training examples? That might be a more common use case (at least for myself).

Yes, that's basically what I have been doing. If I remember correctly, here's what I did – ``` py def custom_objective(...): ... weights_per_label = theano.shared(lasagne.utils.floatX([0.8, 0.1, 0.1])) weights = weights_per_label[target]...

I always find the defaults strange, and I actually changed the defaults in my local clone of Lasagne. Maybe there should be some comments about the need to change the...

As crazy as it sounds, have you guys considered having fewer defaults and requiring the user the define every important parameters? > Where would you look for this? Adding this...

I did not end up using PCA because it worsen the accuracy in my case. But obviously feel free to experiment and report back!