WZMIAOMIAO/deep-learning-for-image-processing · error · ValueError

return_layers are not present in model

Error message

return_layers are not present in model

What it means

BackboneWithFPN's constructor validates that every key of return_layers is an actual named child of the model; if any requested layer name is missing it raises ValueError('return_layers are not present in model'). This prevents silently building an FPN hooked to nonexistent intermediate layers.

Source

Thrown at pytorch_object_detection/faster_rcnn/backbone/feature_pyramid_network.py:34

    This means that one should **not** reuse the same nn.Module
    twice in the forward if you want this to work.
    Additionally, it is only able to query submodules that are directly
    assigned to the model. So if `model` is passed, `model.feature1` can
    be returned, but not `model.feature1.layer2`.
    Arguments:
        model (nn.Module): model on which we will extract the features
        return_layers (Dict[name, new_name]): a dict containing the names
            of the modules for which the activations will be returned as
            the key of the dict, and the value of the dict is the name
            of the returned activation (which the user can specify).
    """
    __annotations__ = {
        "return_layers": Dict[str, str],
    }

    def __init__(self, model, return_layers):
        if not set(return_layers).issubset([name for name, _ in model.named_children()]):
            raise ValueError("return_layers are not present in model")

        orig_return_layers = return_layers
        return_layers = {str(k): str(v) for k, v in return_layers.items()}
        layers = OrderedDict()

        # 遍历模型子模块按顺序存入有序字典
        # 只保存layer4及其之前的结构，舍去之后不用的结构
        for name, module in model.named_children():
            layers[name] = module
            if name in return_layers:
                del return_layers[name]
            if not return_layers:
                break

        super().__init__(layers)
        self.return_layers = orig_return_layers

    def forward(self, x):

View on GitHub (pinned to 1ec3fe6f37)

Solutions

1. Print [name for name,_ in model.named_children()] and use exactly those names as return_layers keys.
2. For ResNet use return_layers = {'layer1':'0','layer2':'1','layer3':'2','layer4':'3'}.
3. Match your backbone to the example the code was written for, or update return_layers for your architecture.
4. Ensure you wrap the raw backbone model once, not an already-wrapped backbone.
5. Check torchvision version for renamed children (e.g. mobilenet_v2 uses 'features', not named stages).

Example fix

# before
return_layers = {'layer1': '0', 'layer2': '1', 'layer3': '2', 'layer5': '3'}  # layer5 doesn't exist
backbone = BackboneWithFPN(resnet50_fpn_backbone, return_layers=return_layers)
# after
return_layers = {'layer1': '0', 'layer2': '1', 'layer3': '2', 'layer4': '3'}
backbone = BackboneWithFPN(resnet50_fpn_backbone, return_layers=return_layers)

children = {name for name, _ in model.named_children()}
assert set(return_layers).issubset(children), \
    f"unknown layers {set(return_layers) - children}; available: {children}"

def layers_exist(model, return_layers: dict) -> bool:
    children = {name for name, _ in model.named_children()}
    return set(return_layers).issubset(children)

try:
    backbone = BackboneWithFPN(model, return_layers=return_layers)
except ValueError as e:
    logging.error("%s — children: %s", e, [n for n, _ in model.named_children()])
    raise SystemExit(1)

When it happens

Trigger: Calling BackboneWithFPN(resnet, return_layers={'layer5':'0'}) or {'conv1':'0'}-style dicts where a key doesn't match model.named_children(); passing a model whose architecture differs (e.g. torchvision version renamed layers, or wrapping an already-FPN-wrapped backbone so expected children vanish).

Common situations: Copy-pasting return_layers from a ResNet-50 example into a different backbone (MobileNet/VGG have different child names); torchvision version changes changing module names; passing layer names instead of child module names (e.g. 'layer1.0.conv1'); applying BackboneWithFPN twice.

Related errors

• backbone should contain an attribute out_channelsspecifying
• illegal stride value.
• The inverted_residual_setting should not be empty.
• The inverted_residual_setting should be List[InvertedResidua
• expected stages_repeats as list of 3 positive ints

AI-assisted analysis of WZMIAOMIAO/deep-learning-for-image-processing@1ec3fe6f37 (2026-08-30). Data as JSON: /api/errors/e3c79f8264c3d577. Report an issue: GitHub.