CNN embedding net fails for small D inputs

[janfb]

the CNN embedding net

https://github.com/mackelab/sbi/blob/a240efb67576608c2cecc7020431ca5baad386a5/sbi/neural_nets/embedding_nets.py#L46-L57

fails for input_dim < 16 and for many other combinations of input parameters.

We should test this in more detail and add some checks on the input params.
And I think we should note that this is a 1D convolutional net, and add a 2D net as well?

[coschroeder]

it is noted that it's a 1d convolution and there is also a remark

        Remark: The implementation of the convolutional layers was not tested
        rigourously. While it works for the default configuration parameters it
        might cause shape conflicts fot badly chosen parameters.

But I agree, that this could be tested more rigorously. That's also why I wanted to fix the channel dimensions in the init and not allow them to be passed as arguments.
If we really want to go down that road, we should also add the flexibility for more layers etc. But not even sure if we can avoid all conflicts by testing the arguments, as there may be all kinds of non-working parameter combinations.

[michaeldeistler]

I think this is fixed @janfb ?

[janfb]

yes, fixed 🎉