Simple easy to use Pytorch module to get the intermediate layers outputs from chosen submodules. Inspired in this but does not assume that submodules are executed sequentially. Pypi link

List of features:

• Supports submodule annidation (module1.submodule2.submodule3)
• In case that a module is called more than once during a forward pass, all it's outputs are saved a in a list.

Installation

pip install torch-intermediate-layer-getter

Usage

Example

import torch
import torch.nn as nn

from torch_intermediate_layer_getter import IntermediateLayerGetter as MidGetter

class Model(nn.Module):
    def __init__(self):
        super().__init__()

        self.fc1 = nn.Linear(2, 2)
        self.fc2 = nn.Linear(2, 2)
        self.nested = nn.Sequential(
            nn.Sequential(nn.Linear(2, 2), nn.Linear(2, 3)),
            nn.Linear(3, 1),
        )
        self.interaction_idty = nn.Identity() # Simple trick for operations not performed as modules

    def forward(self, x):
        x1 = self.fc1(x)
        x2 = self.fc2(x)

        interaction = x1 * x2
        self.interaction_idty(interaction)

        x_out = self.nested(interaction)

        return x_out
        
model = Model()
return_layers = {
    'fc2': 'fc2',
    'nested.0.1': 'nested',
    'interaction_idty': 'interaction',
}
mid_getter = MidGetter(model, return_layers=return_layers, keep_output=True)
mid_outputs, model_output = mid_getter(torch.randn(1, 2))

print(model_output)
>> tensor([[0.3219]], grad_fn=<AddmmBackward>)
print(mid_outputs)
>> OrderedDict([('fc2', tensor([[-1.5125,  0.9334]], grad_fn=<AddmmBackward>)),
  ('interaction', tensor([[-0.0687, -0.1462]], grad_fn=<MulBackward0>)),
  ('nested', tensor([[-0.1697,  0.1432,  0.2959]], grad_fn=<AddmmBackward>))])

# model_output is None if keep_ouput is False
# if keep_output is True the model_output contains the final model's output