dacapo.experiments.model

Classes

Model

A trainable DaCapo model. Consists of an Architecture and a

Module Contents

class dacapo.experiments.model.Model(architecture: dacapo.experiments.architectures.architecture.Architecture, prediction_head: torch.nn.Module, eval_activation: torch.nn.Module | None = None)

A trainable DaCapo model. Consists of an Architecture and a prediction head. Models are generated by ``Predictor``s.

May include an optional eval_activation that is only executed when the model is in eval mode. This is particularly useful if you want to train with something like BCELossWithLogits, since you want to avoid applying softmax while training, but apply it during evaluation.

architecture

The architecture of the model.

Type:: Architecture

prediction_head

The prediction head of the model.

Type:: torch.nn.Module

chain

The architecture followed by the prediction head.

Type:: torch.nn.Sequential

num_in_channels

The number of input channels.

Type:: int

input_shape

The shape of the input tensor.

Type:: Coordinate

eval_input_shape

The shape of the input tensor during evaluation.

Type:: Coordinate

num_out_channels

The number of output channels.

Type:: int

output_shape

The shape of the output

Type:: Coordinate

eval_activation

The activation function to apply during evaluation.

Type:: torch.nn.Module | None

forward(x: torch.Tensor) -> torch.Tensor: Forward pass of the model.

compute_output_shape(input_shape: Coordinate) -> Tuple[int, Coordinate]: Compute the spatial shape of this model, when fed a tensor of the given spatial shape as input.

scale(voxel_size: Coordinate) -> Coordinate: Scale the model by the given voxel size.

Note

The output shape is the spatial shape of the model, i.e., not accounting for channels and batch dimensions.

num_out_channels: int

num_in_channels: int

architecture

prediction_head

chain

input_shape

eval_input_shape

eval_activation

forward(x)

Forward pass of the model.

Parameters:: x (torch.Tensor) – The input tensor.
Returns:: The output tensor.
Return type:: torch.Tensor

Examples

>>> model = Model(architecture, prediction_head)
>>> model.forward(x)
torch.Tensor

Note

The eval_activation is only applied during evaluation. This is particularly useful if you want to train with something like BCELossWithLogits, since you want to avoid applying softmax while training, but apply it during evaluation.

compute_output_shape(input_shape: funlib.geometry.Coordinate) → Tuple[int, funlib.geometry.Coordinate]

Compute the spatial shape (i.e., not accounting for channels and batch dimensions) of this model, when fed a tensor of the given spatial shape as input.

Parameters:: input_shape (Coordinate) – The shape of the input tensor.
Returns:: The number of output channels and the spatial shape of the output.
Return type:: Tuple[int, Coordinate]
Raises:: AssertionError – If the input_shape is not a Coordinate.

Examples

>>> model = Model(architecture, prediction_head)
>>> model.compute_output_shape(input_shape)
(1, Coordinate(1, 1, 1))

Note

The output shape is the spatial shape of the model, i.e., not accounting for channels and batch dimensions.

scale(voxel_size: funlib.geometry.Coordinate) → funlib.geometry.Coordinate

Scale the model by the given voxel size.

Parameters:: voxel_size (Coordinate) – The voxel size to scale the model by.
Returns:: The scaled model.
Return type:: Coordinate
Raises:: AssertionError – If the voxel_size is not a Coordinate.

Examples

>>> model = Model(architecture, prediction_head)
>>> model.scale(voxel_size)
Coordinate(1, 1, 1)

Note

The output shape is the spatial shape of the model, i.e., not accounting for channels and batch dimensions.