No softmax in densenet

[grudloff]

I was checking the implementation of densenet and I noticed that the class layers, defined as:

  self.class_layers = nn.Sequential(
      OrderedDict(
          [
              ("relu", get_act_layer(name=act)),
              ("pool", avg_pool_type(1)),
              ("flatten", nn.Flatten(1)),
              ("out", nn.Linear(in_channels, out_channels)),
          ]

There is no softmax in the output, and I find this odd. Am I missing something?

[KumoLiu]

Hi @grudloff, I think you can add it by using Activations.
https://github.com/Project-MONAI/MONAI/blob/281cb0119c01eaa8e6c841880b91f92f45e8d7f7/monai/transforms/post/array.py#L59

Thanks!

[ericspod]

We have typically separated network outputs from activation so that you have the flexibility to do what you like with those raw values. Activation is done with postprocessing transforms like this:

tutorials/3d_classification/ignite/densenet_training_dict.py

Line 124 in 02a1664

post_pred = Compose([Activations(softmax=True)])