Feature extraction for model inspection

The torchvision.models.feature_extraction package contains feature extraction utilities that let us tap into our models to access intermediate transformations of our inputs. This could be useful for a variety of applications in computer vision. Just a few examples are:

• Visualizing feature maps.

• Extracting features to compute image descriptors for tasks like facial recognition, copy-detection, or image retrieval.

• Passing selected features to downstream sub-networks for end-to-end training with a specific task in mind. For example, passing a hierarchy of features to a Feature Pyramid Network with object detection heads.

Torchvision provides create_feature_extractor() for this purpose. It works by following roughly these steps:

1. Symbolically tracing the model to get a graphical representation of how it transforms the input, step by step.

2. Setting the user-selected graph nodes as outputs.

3. Removing all redundant nodes (anything downstream of the output nodes).

4. Generating python code from the resulting graph and bundling that into a PyTorch module together with the graph itself.

The torch.fx documentation provides a more general and detailed explanation of the above procedure and the inner workings of the symbolic tracing.

About Node Names

In order to specify which nodes should be output nodes for extracted features, one should be familiar with the node naming convention used here (which differs slightly from that used in torch.fx). A node name is specified as a . separated path walking the module hierarchy from top level module down to leaf operation or leaf module. For instance "layer4.2.relu" in ResNet-50 represents the output of the ReLU of the 2nd block of the 4th layer of the ResNet module. Here are some finer points to keep in mind:

• When specifying node names for create_feature_extractor(), you may provide a truncated version of a node name as a shortcut. To see how this works, try creating a ResNet-50 model and printing the node names with train_nodes, _ = get_graph_node_names(model) print(train_nodes) and observe that the last node pertaining to layer4 is "layer4.2.relu_2". One may specify "layer4.2.relu_2" as the return node, or just "layer4" as this, by convention, refers to the last node (in order of execution) of layer4.

• If a certain module or operation is repeated more than once, node names get an additional _{int} postfix to disambiguate. For instance, maybe the addition (+) operation is used three times in the same forward method. Then there would be "path.to.module.add", "path.to.module.add_1", "path.to.module.add_2". The counter is maintained within the scope of the direct parent. So in ResNet-50 there is a "layer4.1.add" and a "layer4.2.add". Because the addition operations reside in different blocks, there is no need for a postfix to disambiguate.

An Example

Here is an example of how we might extract features for MaskRCNN:

import torch
from torchvision.models import resnet50
from torchvision.models.feature_extraction import get_graph_node_names
from torchvision.models.feature_extraction import create_feature_extractor
from torchvision.models.detection.mask_rcnn import MaskRCNN
from torchvision.models.detection.backbone_utils import LastLevelMaxPool
from torchvision.ops.feature_pyramid_network import FeaturePyramidNetwork


# To assist you in designing the feature extractor you may want to print out
# the available nodes for resnet50.
m = resnet50()
train_nodes, eval_nodes = get_graph_node_names(resnet50())

# The lists returned, are the names of all the graph nodes (in order of
# execution) for the input model traced in train mode and in eval mode
# respectively. You'll find that `train_nodes` and `eval_nodes` are the same
# for this example. But if the model contains control flow that's dependent
# on the training mode, they may be different.

# To specify the nodes you want to extract, you could select the final node
# that appears in each of the main layers:
return_nodes = {
    # node_name: user-specified key for output dict
    'layer1.2.relu_2': 'layer1',
    'layer2.3.relu_2': 'layer2',
    'layer3.5.relu_2': 'layer3',
    'layer4.2.relu_2': 'layer4',
}

# But `create_feature_extractor` can also accept truncated node specifications
# like "layer1", as it will just pick the last node that's a descendent of
# of the specification. (Tip: be careful with this, especially when a layer
# has multiple outputs. It's not always guaranteed that the last operation
# performed is the one that corresponds to the output you desire. You should
# consult the source code for the input model to confirm.)
return_nodes = {
    'layer1': 'layer1',
    'layer2': 'layer2',
    'layer3': 'layer3',
    'layer4': 'layer4',
}

# Now you can build the feature extractor. This returns a module whose forward
# method returns a dictionary like:
# {
#     'layer1': output of layer 1,
#     'layer2': output of layer 2,
#     'layer3': output of layer 3,
#     'layer4': output of layer 4,
# }
create_feature_extractor(m, return_nodes=return_nodes)

# Let's put all that together to wrap resnet50 with MaskRCNN

# MaskRCNN requires a backbone with an attached FPN
class Resnet50WithFPN(torch.nn.Module):
    def __init__(self):
        super(Resnet50WithFPN, self).__init__()
        # Get a resnet50 backbone
        m = resnet50()
        # Extract 4 main layers (note: MaskRCNN needs this particular name
        # mapping for return nodes)
        self.body = create_feature_extractor(
            m, return_nodes={f'layer{k}': str(v)
                             for v, k in enumerate([1, 2, 3, 4])})
        # Dry run to get number of channels for FPN
        inp = torch.randn(2, 3, 224, 224)
        with torch.no_grad():
            out = self.body(inp)
        in_channels_list = [o.shape[1] for o in out.values()]
        # Build FPN
        self.out_channels = 256
        self.fpn = FeaturePyramidNetwork(
            in_channels_list, out_channels=self.out_channels,
            extra_blocks=LastLevelMaxPool())

    def forward(self, x):
        x = self.body(x)
        x = self.fpn(x)
        return x


# Now we can build our model!
model = MaskRCNN(Resnet50WithFPN(), num_classes=91).eval()

API Reference

create_feature_extractor(model[, …])	Creates a new graph module that returns intermediate nodes from a given model as dictionary with user specified keys as strings, and the requested outputs as values.
get_graph_node_names(model[, tracer_kwargs, …])	Dev utility to return node names in order of execution.