Named layers in tensorboard graph of dynamic Module

When I directly define a nn.Module with fixed layers and use torch.utils.tensorboard.SummaryWriter, I get a nice vizualisation of my computational graph like this::

class NeuralNetStrategy(nn.Module):
    def __init__(self, input_length):
        super(NeuralNetStrategy, self).__init__()
        self.fc1 = nn.Linear(input_length, 10) 
        self.fc_out = nn.Linear(10, 1)
    def forward(self, x):
        x = self.fc1(x).tanh()
        x = self.fc_out(x).relu()
        return x

model = NeuralNetStrategy(1)
with SummaryWriter(logdir) as writer:

My goal is to get a similar output graph but for a dynamic number of layers created at runtime, but I’m having trouble to make this work.

I’ve tried using nn.ModuleDict() or building an OrderedDict and passing it to nn.Sequential in order to create named layers, but my tensorboard graph does not display the names. My current class definition and tensorboard output looks something like this:

class NeuralNetStrategy(nn.Module):
    def __init__(self, input_length, hidden_nodes: Iterable[int], hidden_activations: Iterable):
        super(NeuralNetStrategy, self).__init__()
        self.layers = nn.ModuleDict() 
        self.layers['fc_0'] = nn.Linear(input_length, hidden_nodes[0])
        self.layers['activation_0'] = hidden_activations[0]

        for i in range (1, len(hidden_nodes)):
            self.layers['fc_' + str(i)] = nn.Linear(hidden_nodes[i-1], hidden_nodes[i])
            self.layers['activation_' + str(i)] = hidden_activations[i]

        self.layers['fc_out'] = nn.Linear(hidden_nodes[-1], 1)
        self.layers['activation_out'] = nn.ReLU()
    def forward(self, x):
        for layer in self.layers.values():
            x = layer(x)
        return x

When I initialize my model, it looks fine in pytorch (and does perform the desired computations), yet the tensorboard graph does not respect the named layers:



How can I fix this?

1 Like