Mikaël Jacquemont

Hi @Jacobew and @tamaraalshekhli , the task noise is an additional trainable parameter, although not an additional output of your network. What makes it the noise is the computation of...

@tamaraalshekhli , if you use TF, it should be straightforward as every thing is described in Y. Gal's notebook.

@tamaraalshekhli I'm not familiar with Keras (I use PyTorch), but if I understand correctly, get_prediction_model defines the model that learns to reconstruct the data and get_trainable_model combines the output of...

Hi @TjarkMiener, to use indexed convolution (or pooling) you also need at least `neighbours_extraction` and `pool_index_matrix`. It would be very nice if you add the tf version directly in the...

@TjarkMiener, I've prepared IndexedConv for tf in the branch tf_support.

@TjarkMiener , could you please provide the error trace you get when trying to install ?

@TjarkMiener , I've fixed the installation and import issues (112bbe9).

Investigating on why indexedconv uses a lot more memory than built-in conv and is quite slower I found: * https://discuss.pytorch.org/t/matmul-broadcasting-makes-copies/19494 * https://discuss.pytorch.org/t/memory-inefficient-in-batch-matrix-multiplication-with-autograd/28164 I'm doing some tests to confirm that it...

The memory consumption observed with indexedconv is due to the matmul function which operates broadcasting on tensors before applying matrix multiplication. Indeed the weight matrix needs to be expanded to...

> hi @mikael10j. > Does it solve entirely the memory usage then? > (you might still have some small overhead with Python compared to the cuda version) > > I...