Jan Schlüter

Recurrent layers discard updates dictionary returned by scan

16

From the mailing list: https://groups.google.com/forum/#!topic/lasagne-users/8E1iR04_nK8 The following code fails: ``` python import theano import lasagne n_batch, n_steps, n_in = (2, 3, 4) n_hid = 5 l_in = lasagne.layers.InputLayer((n_batch, n_steps, n_in))...

InverseLayer cannot invert MergeLayer

1

The `InverseLayer` currently assumes the layer to be inverted has a single incoming layer: https://github.com/Lasagne/Lasagne/blob/b6c3e5c/lasagne/layers/special.py#L344 The `InverseLayer` should get an optional third argument specifying the input to backpropagate to. This...

API: Multiple outputs in Layer.get_output_for()

32

In the context of recurrent neural networks, we've found two use cases that would need layers to be able to return multiple outputs: 1. @skaae wanted to return the hidden...

Provide RNN "containers"

84

In https://github.com/Lasagne/Recipes/pull/23#issuecomment-141979978, @skaae brought up the idea of splitting up the recurrent layers into layers performing a single computation step and a container layer that handles the recurrence (i.e., sets...

MergeLayer accepting dictionaries

6

Right now, our MergeLayer base class, the only layer class that can handle multiple inputs, requires its input layers to be given as a list. This is a good fit...

Add cyclic pooling and rolling layers?

3

Would be nice to support cyclic pooling and cyclic rolling in Lasagne out-of-the-box. It has been shown useful for many tasks (Galaxy classification, Plankton classification, and probably also for playing...

First of all, thanks a lot for packaging this! I've noticed some interesting errors when creating a pairwise potential with very small stddev for the feature comparison. Minimal example: ```python...

Just thought we could collect some ideas for things that'd be reasonably easy to turn into Lasagne Recipes. One of them is Ben Graham's "Index-learning of unsupervised low dimensional embeddings",...

Simplify setup.py with custom compiler class

2

This removes the `build_ext` hack overriding `compiler.spawn` to modify the compilation commands into proper `nvcc` calls depending on the platform (i.e., do nothing on Unix, but a lot on Windows)....

Compile for higher GPU architectures

6

Changing the Makefile to match the compute capabilities of the device at hands can provide performance benefits. For example, on a GTX 580 (compute capability 2.0), adding `-arch=sm_20` to the...