Confused with ".act = nn.Identity()"

1

I am trying to understand a model implementation where there is this following layer. I don’t understand what the last line means.

class Model(nn.Module):
    def __init__(self,
                 num_class,
                 num_point,
                 num_person,
                 num_gcn_scales,
                 num_g3d_scales,
                 graph,
                 in_channels=3):
        super(Model, self).__init__()

        Graph = import_class(graph)
        A_binary = Graph().A_binary
        
        self.sgcn2 = nn.Sequential(
        MS_GCN(num_gcn_scales, c1, c1, A_binary, disentangled_agg=True),
        MS_TCN(c1, c2, stride=2),
        MS_TCN(c2, c2))
        self.sgcn2[-1].act = nn.Identity()

It looks like the model first perform spatial graph convolution followed by two temporal convolution. But what happens with this line self.sgcn2[-1].act = nn.Identity()

Your suggestions will greatly help. Thanks.

2

Hi,

self.sgcn2 is sequential of MS_GCN, MS_TCN, and MS_TCN.
So, self.sgcn2[-1] refers to the last MS_TCN.
self.sgcn2[-1].act is either the attribute or member variable of an instance of MS_TCN. From that name, I guess by default MS_TCN has some activation function such as Sigmoid and ReLU. Though the author of that snippet wanted that last module to use Identity as its activation.

3

Thank you so much. That is exactly the case, there is a activation function in MS_TCN and act holds the value that determines what the activation will be.