capsa_torch.neo

class capsa_torch.neo.Wrapper
__init__(integration_sites=2, layer_alpha=(2.0, 1.0), layer_out_dims=None, finetune=False, conv_params=None, add_batch_norm=False, node_name_filter=None, pixel_wise=False, *, verbose=0, torch_compile=False, symbolic_trace=False)

Initialize a Neo Wrapper with configs. :type integration_sites: int :param integration_sites: The number of sites to use when integrating neo wrapper into your model. More integration sites may produce more robust vacuitic uncertainty estimates but will increase computation costs. (default: 2) :type layer_alpha: tuple[float, …] :param layer_alpha: Controls the structure of neo integrations. Pair of floats. Larger values produce more robust vacuitic uncertainty estimates but with more compute and memory overhead. (default: (2.0, 1.0)) :type layer_out_dims: tuple[int, …] | None :param layer_out_dims: Optional sizes for the output dimensions of integration layers. If provided, must be a tuple matching the number of integration sites. (default: None) :type finetune: bool :param finetune: Freeze the existing module weights and only train the added parameters. (default: False) :type conv_params: dict | None :param conv_params: Dictionary of convolution parameters for integration layers. Supported keys:

  • kernel_size: size of the convolution kernels (int)

  • padding: padding applied to the layers (int)

  • stride: stride of the convolution (int) (default: None)

Parameters:

  • add_batch_norm (bool) – Whether to insert BatchNorm layers before the integration convolutions. (default: False)

  • node_name_filter (list[str] | None) – List of node names to filter out from the graph. If provided, must be a list matching the number of integration sites, or a list of a single string. (default: None)

  • pixel_wise (bool) – If True, compute vacuity scores per pixel; otherwise, compute a global score. (default: False)

  • verbose (int) – Set the verbosity level for wrapping. 0 <= verbose <= 2 (default: 0)

  • torch_compile (bool) – Apply torch’s inductor to compile the wrapped model. This should improve model performance, at the cost of initial overhead. (default: False)

  • symbolic_trace (bool) – Attempt to use symbolic shapes when tracing the module’s graph. Turning this on will help to avoid rewrapping when the model is fed different input shapes, however it is not well supported by all models. (default: False)

Note

verbose and symbolic_trace are keyword arguments only

__call__(module_or_module_class)

Applies wrapper to either an instantiated torch.nn.Module or a class that subclasses torch.nn.Module to create a new wrapped implementation.

Parameters:

module_or_module_class (TypeVar(T, bound= torch.nn.Module | type[torch.nn.Module])) – The Module to wrap

Return type:

TypeVar(T, bound= torch.nn.Module | type[torch.nn.Module])

Returns:

The wrapped module, with weights shared with module

Example Usage

Wrapping a Module
from capsa_torch.sample import Wrapper # or capsa_torch.sculpt, capsa_torch.vote
wrapper = Wrapper(n_samples=3, verbose=1) # Initialize a wrapper object with your config options

wrapped_module = wrapper(module) # wrap your module

y = wrapped_module(x) # Use the wrapped module as usual
y, risk = wrapped_module(x, return_risk=True) # Use the wrapped module to obtain risk values
Decorator approach
from capsa_torch.sample import Wrapper # or capsa_torch.sculpt, capsa_torch.vote

@Wrapper(n_samples=3, verbose=1) # Initialize a wrapper object with your config options
class MyModule(torch.nn.Module): # Note: MyModule must subclass torch.nn.Module
    def __init__(self, ...):
        ...

    def forward(self, ...):
        ...

wrapped_module = MyModule(...) # Call MyModule's __init__ fn as usual to create a wrapped module

y = wrapped_module(x) # Use the wrapped module as usual
y, risk = wrapped_module(x, return_risk=True) # Use the wrapped module to obtain risk values