Creating an embedding net with batchnorm 1D fails due to check of embedding net outputsize

[rmehta1987]

When initializing and building the initial flow, these statements can raise exceptions, https://github.com/mackelab/sbi/blob/0fcf9842c9515fc089ea464fb11359d4c19c7331/sbi/neural_nets/flow.py#L59

When an embedding net with batchnorm1D is used, batchnorm expects a batch of inputs. Is it not better to see the output size of the layer by doing:

embedding_net.model[-1].out_features ?

[manuelgloeckler]

Hey,

Thanks for reporting this limitation. Your suggestion however won't work in general as not all PyTorch Modules have the attribute out_features i.e. activation functions.

There should be two ways to fix this:

• Giving the embedding net a batch of at least two datapoints.
• Switch to evaluation mode for this operation i.e. emebdding_net.eval().

If you pass the embedding net in evaluation mode to the flow builder, it should also work, no? Later during training it will automatically be switched to train.

Kind regards, Manuel

[rmehta1987]

Hi Manuel,

Your suggestions also work. I just wanted to point out that this happens during the initialization phase, density_estimator_function = posterior_nn(model="maf", embedding_net=embedding_net, hidden_features=num_hidden, num_transforms=number_of_transforms) which then calls build_maf, which then does y_numel = embedding_net(batch_y[:1]).numel().

[felixp8]

I can try to take this