Define layers as dict

1

Hi,

I have the following code:

class LeNet5(nn.Module):

	def __init__(self, num_classes, grayscale=False):
		super(LeNet5, self).__init__()
		
		self.grayscale = grayscale
		self.num_classes = num_classes

		if self.grayscale:
			in_channels = 1
		else:
			in_channels = 3

		# features layers
		self.layer = {}
		self.layer[1] = nn.Sequential(
			nn.Conv2d(in_channels, 6, kernel_size=5),
			nn.Tanh()
		)
		self.layer[2] = nn.Sequential(
			nn.MaxPool2d(kernel_size=2)
		)
		self.layer[3] = nn.Sequential(
			nn.Conv2d(6, 16, kernel_size=5),
			nn.Tanh()
		)
		self.layer[4] = nn.Sequential(
			nn.MaxPool2d(kernel_size=2)
		)

		# classifier, linears layers
		self.layer[5] = nn.Sequential(
			nn.Linear(16*5*5, 120),
			nn.Tanh()
		)
		self.layer[6] = nn.Sequential(
			nn.Linear(120, 84),
			nn.Tanh()
		)
		self.layer[7] = nn.Sequential(
			nn.Linear(84, num_classes)
		)

	def forward(self, x):
		forward_layer = {}

		forward_layer[0] = x
		forward_layer[1] = self.layer[1](forward_layer[0])
		forward_layer[2] = self.layer[2](forward_layer[1])
		forward_layer[3] = self.layer[3](forward_layer[2])
		forward_layer[4] = self.layer[4](forward_layer[3])

		forward_layer[4] = torch.flatten(forward_layer[4], 1)
		forward_layer[5] = self.layer[5](forward_layer[4])
		forward_layer[6] = self.layer[6](forward_layer[5])
		forward_layer[7] = self.layer[7](forward_layer[6])

		logits = forward_layer[7]
		probas = F.softmax(logits, dim=1)
		return logits, probas, x, forward_layer

It returns the error:

ValueError: optimizer got an empty parameter list

And I verified that the model is empty.

If I change the nams of the layers inside the class it works.

Can I define, inside the class, the layers as dict?

2

Use nn.ModuleDict instead of a plain Python dict to properly register the modules.

1 Like
3

It doesn’t accept integer number as keys, right?

4

I wanted to access the wieghts of the layer dynamically. After using nn.ModuleDict I will share the solution. If someone have some better solution I ask to share. Thanks

for s in model.layers.items():
    for l in s[1].parameters():
        print(l)