visualize model that not in the model zoo about pytorch-cnn-visualizations

[utkuozbulak]

Hey, it doesn't have to contain features, what is important is for the model to contain nn.Sequential so that you can iterate.

Let me elaborate.

You can have an __ init __() function like this:

...
def __init__(self):
  self.my_layers = nn.Sequential(
    nn.Conv2d(1, 6, (5, 5), padding=2),
    nn.Conv2d(1, 6, (5, 5), padding=2),
    nn.Conv2d(1, 6, (5, 5), padding=2),
    .
    .
    .
 )

def forward(self, input):
  output = self.my_layers(input)

or this

...
def __init__(self):
    self.layer1 = nn.Conv2d(1, 6, (5, 5), padding=2)
    self.layer2 = nn.Conv2d(1, 6, (5, 5), padding=2)
    self.layer3 = nn.Conv2d(1, 6, (5, 5), padding=2)
    .
    .
    .
 
def forward(self, x):
  x = self.layer1(x)
  x = self.layer2(x)
  x = self.layer3(x)
  ...

If you have a model that wraps up layers with nn.Sequential then iterating through layers one by one is super easy.

x = input
for single_layer in model.my_layers():
   x = single_layer(x)

If there is no nn.Sequential wrapping it up, you can't use this form (becase there is no pre-defined iterative process) but you can use .modules() or .named_children().

Have a look at these:
https://discuss.pytorch.org/t/module-children-vs-module-modules/4551
https://discuss.pytorch.org/t/how-to-loop-over-all-the-variable-in-a-nn-module/912

Hope it helps.

[Raunak13]

Hi @visonpon , @utkuozbulak , I am getting the same errors. Can you tell me what did you do?
While using model.getnamedchildren() the function 'get_output_from_specific_layer' returns values for first few layers, but then it gives the error: "RuntimeError: Given groups=1, weight[48, 192, 1, 1], so expected input[1, 64, 105, 105] to have 192 channels, but got 64 channels instead"
I am using Inception Resnet V3

[utkuozbulak]

I'm also getting some emails about how hard the whole procedure becomes when the model has nested elements (like Resnets). I might just make a ResNet example soon.

[himat]

Has a ResNet example been made?

[Saharkakavand]

@utkuozbulak did you make that example about resnet?

[utkuozbulak]

I had one done for one of my projects but did not put it in this repository.

Adapting the code to resnet, or for any other architecture that has nested blocks, is not hard. You only need to change two or three lines of code where it hooks the layers. Because resnet has residual blocks, what you want is to hook the layers inside these blocks and not the blocks themselves.