sleap.nn.config.model

class sleap.nn.config.model.BackboneConfig(*args, **kwargs)[source]

Configurations related to the model backbone.

Only one field can be set and will determine which backbone architecture to use.

leap

A LEAPConfig instance.

Type

Optional[sleap.nn.config.model.LEAPConfig]

unet

A UNetConfig instance.

Type

Optional[sleap.nn.config.model.UNetConfig]

hourglass

A HourglassConfig instance.

Type

Optional[sleap.nn.config.model.HourglassConfig]

resnet

A ResNetConfig instance.

Type

Optional[sleap.nn.config.model.ResNetConfig]

class sleap.nn.config.model.CenteredInstanceConfmapsHeadConfig(anchor_part: Optional[str] = None, part_names: Optional[List[str]] = None, sigma: float = 5.0, output_stride: int = 1, offset_refinement: bool = False)[source]

Configurations for centered instance confidence map heads.

These heads are used in topdown multi-instance models that make the assumption that there is an instance reliably centered in the cropped input image. These heads are useful when centroids are easy to detect as they learn complex relationships between the geometry of body parts, even when animals are occluded.

This comes at the cost of a strong reliance on the accuracy of the instance-centered cropping, i.e., it is heavily limited by the accuracy of the centroid model.

Additionally, since one image crop is evaluated per instance, topdown models scale linearly with the number of animals in the frame, which can result in poor performance when many instances are present.

Use this head when centroids are easy to detect, preferably using a consistent body part as an anchor, and when there are few animals that cover a small region of the full frame.

anchor_part

Text name of a body part (node) to use as the anchor point. If None, the midpoint of the bounding box of all visible instance points will be used as the anchor. The bounding box midpoint will also be used if the anchor part is specified but not visible in the instance. Setting a reliable anchor point can significantly improve topdown model accuracy as they benefit from a consistent geometry of the body parts relative to the center of the image.

Type

Optional[str]

part_names

Text name of the body parts (nodes) that the head will be configured to produce. The number of parts determines the number of channels in the output. If not specified, all body parts in the skeleton will be used.

Type

Optional[List[str]]

sigma

Spread of the Gaussian distribution of the confidence maps as a scalar float. Smaller values are more precise but may be difficult to learn as they have a lower density within the image space. Larger values are easier to learn but are less precise with respect to the peak coordinate. This spread is in units of pixels of the model input image, i.e., the image resolution after any input scaling is applied.

Type

float

output_stride

The stride of the output confidence maps relative to the input image. This is the reciprocal of the resolution, e.g., an output stride of 2 results in confidence maps that are 0.5x the size of the input. Increasing this value can considerably speed up model performance and decrease memory requirements, at the cost of decreased spatial resolution.

Type

int

offset_refinement

If True, model will also output an offset refinement map used to achieve subpixel localization of peaks during inference. This can improve the localization accuracy of the model at the cost of additional memory and training and inference time. If False (the default), subpixel localization can be achieved post-hoc with deterministic refinement, which does not require additional resources or training, but may not achieve the same accuracy as learned refinement.

Type

bool

class sleap.nn.config.model.CentroidsHeadConfig(anchor_part: Optional[str] = None, sigma: float = 5.0, output_stride: int = 1, offset_refinement: bool = False)[source]

Configurations for centroid confidence map heads.

These heads are used in topdown models that rely on centroid detection to detect instances for cropping before predicting the remaining body parts.

Multiple centroids can be present (one per instance), so their coordinates can be recovered in inference via local peak finding.

anchor_part

Text name of a body part (node) to use as the anchor point. If None, the midpoint of the bounding box of all visible instance points will be used as the anchor. The bounding box midpoint will also be used if the anchor part is specified but not visible in the instance. Setting a reliable anchor point can significantly improve topdown model accuracy as they benefit from a consistent geometry of the body parts relative to the center of the image.

Type

Optional[str]

sigma

Spread of the Gaussian distribution of the confidence maps as a scalar float. Smaller values are more precise but may be difficult to learn as they have a lower density within the image space. Larger values are easier to learn but are less precise with respect to the peak coordinate. This spread is in units of pixels of the model input image, i.e., the image resolution after any input scaling is applied.

Type

float

output_stride

The stride of the output confidence maps relative to the input image. This is the reciprocal of the resolution, e.g., an output stride of 2 results in confidence maps that are 0.5x the size of the input. Increasing this value can considerably speed up model performance and decrease memory requirements, at the cost of decreased spatial resolution.

Type

int

offset_refinement

If True, model will also output an offset refinement map used to achieve subpixel localization of peaks during inference. This can improve the localization accuracy of the model at the cost of additional memory and training and inference time. If False (the default), subpixel localization can be achieved post-hoc with deterministic refinement, which does not require additional resources or training, but may not achieve the same accuracy as learned refinement.

Type

bool

class sleap.nn.config.model.HeadsConfig(*args, **kwargs)[source]

Configurations related to the model output head type.

Only one attribute of this class can be set, which defines the model output type.

single_instance

An instance of SingleInstanceConfmapsHeadConfig.

Type

Optional[sleap.nn.config.model.SingleInstanceConfmapsHeadConfig]

centroid

An instance of CentroidsHeadConfig.

Type

Optional[sleap.nn.config.model.CentroidsHeadConfig]

centered_instance

An instance of CenteredInstanceConfmapsHeadConfig.

Type

Optional[sleap.nn.config.model.CenteredInstanceConfmapsHeadConfig]

multi_instance

An instance of MultiInstanceConfig.

Type

Optional[sleap.nn.config.model.MultiInstanceConfig]

class sleap.nn.config.model.HourglassConfig(stem_stride: int = 4, max_stride: int = 64, output_stride: int = 4, stem_filters: int = 128, filters: int = 256, filter_increase: int = 128, stacks: int = 3)[source]

Hourglass backbone configuration.

stem_stride

Controls how many stem blocks to use for initial downsampling. These are useful for learned downsampling that is able to retain spatial information while reducing large input image sizes.

Type

int

max_stride

Determines the number of downsampling blocks in the network, increasing receptive field size at the cost of network size.

Type

int

output_stride

Determines the number of upsampling blocks in the network.

Type

int

filters

Base number of filters in the network.

Type

int

filters_increase

Constant to increase the number of filters by at each block.

stacks

Number of repeated stacks of the network (excluding the stem).

Type

int

class sleap.nn.config.model.LEAPConfig(max_stride: int = 8, output_stride: int = 1, filters: int = 64, filters_rate: float = 2, up_interpolate: bool = False, stacks: int = 1)[source]

LEAP backbone configuration.

max_stride

Determines the number of downsampling blocks in the network, increasing receptive field size at the cost of network size.

Type

int

output_stride

Determines the number of upsampling blocks in the network.

Type

int

filters

Base number of filters in the network.

Type

int

filters_rate

Factor to scale the number of filters by at each block.

Type

float

up_interpolate

If True, use bilinear upsampling instead of transposed convolutions for upsampling. This can save computations but may lower overall accuracy.

Type

bool

stacks

Number of repeated stacks of the network (excluding the stem).

Type

int

class sleap.nn.config.model.ModelConfig(backbone: sleap.nn.config.model.BackboneConfig = NOTHING, heads: sleap.nn.config.model.HeadsConfig = NOTHING)[source]

Configurations related to model architecture.

backbone

Configurations related to the main network architecture.

Type

sleap.nn.config.model.BackboneConfig

heads

Configurations related to the output heads.

Type

sleap.nn.config.model.HeadsConfig

class sleap.nn.config.model.MultiInstanceConfig(confmaps: sleap.nn.config.model.MultiInstanceConfmapsHeadConfig = NOTHING, pafs: sleap.nn.config.model.PartAffinityFieldsHeadConfig = NOTHING)[source]

Configuration for combined multi-instance confidence map and PAF model heads.

This configuration specifies a multi-head model that outputs both multi-instance confidence maps and part affinity fields, which together enable multi-instance pose estimation in a bottom-up fashion, i.e., no instance cropping or centroids are required.

confmaps

Part confidence map configuration (see the description in MultiInstanceConfmapsHeadConfig).

Type

sleap.nn.config.model.MultiInstanceConfmapsHeadConfig

pafs

Part affinity fields configuration (see the description in PartAffinityFieldsHeadConfig).

Type

sleap.nn.config.model.PartAffinityFieldsHeadConfig

class sleap.nn.config.model.MultiInstanceConfmapsHeadConfig(part_names: Optional[List[str]] = None, sigma: float = 5.0, output_stride: int = 1, loss_weight: float = 1.0, offset_refinement: bool = False)[source]

Configurations for multi-instance confidence map heads.

These heads are used in bottom-up multi-instance models that do not make any assumption about the connectivity of the body parts. These heads will generate multiple local peaks for each body part type and must be detected using local peak finding.

Although this head alone is sufficient to detect multiple copies of each body part type, it provides no information as to which sets of points should be grouped together to the same instance. If this is required, a head that provides connectivity or grouping information is required, e.g., part affinity fields.

Use this head when multiple instances of each body part are present and do not need to be grouped or will be grouped using additional information.

This head type has the advantage that it only needs to evaluate each frame once to find all peaks, in contrast to topdown models that must be evaluated for each crop. This constant scaling with the number of instances can be especially beneficial when there are many animals present in the frame.

part_names

Text name of the body parts (nodes) that the head will be configured to produce. The number of parts determines the number of channels in the output. If not specified, all body parts in the skeleton will be used.

Type

Optional[List[str]]

sigma

Spread of the Gaussian distribution of the confidence maps as a scalar float. Smaller values are more precise but may be difficult to learn as they have a lower density within the image space. Larger values are easier to learn but are less precise with respect to the peak coordinate. This spread is in units of pixels of the model input image, i.e., the image resolution after any input scaling is applied.

Type

float

output_stride

The stride of the output confidence maps relative to the input image. This is the reciprocal of the resolution, e.g., an output stride of 2 results in confidence maps that are 0.5x the size of the input. Increasing this value can considerably speed up model performance and decrease memory requirements, at the cost of decreased spatial resolution.

Type

int

loss_weight

Scalar float used to weigh the loss term for this head during training. Increase this to encourage the optimization to focus on improving this specific output in multi-head models.

Type

float

offset_refinement

If True, model will also output an offset refinement map used to achieve subpixel localization of peaks during inference. This can improve the localization accuracy of the model at the cost of additional memory and training and inference time. If False (the default), subpixel localization can be achieved post-hoc with deterministic refinement, which does not require additional resources or training, but may not achieve the same accuracy as learned refinement.

Type

bool

class sleap.nn.config.model.PartAffinityFieldsHeadConfig(edges: Optional[Sequence[Tuple[str, str]]] = None, sigma: float = 15.0, output_stride: int = 1, loss_weight: float = 1.0)[source]

Configurations for multi-instance part affinity field heads.

These heads are used in bottom-up multi-instance models that require information about body part connectivity in order to group multiple detections of each body part type into distinct instances.

Part affinity fields are an image-space representation of the directed graph that defines the skeleton. Pixels that are close to the line (directed edge) formed between pairs of nodes of the same instance will contain unit vectors pointing along the direction of the the connection. The similarity between this line and the average of the unit vectors at the pixels underneath the line can be used as a matching score to associate candidate pairs of body part detections.

Use this head when multiple instances of each body part are present and need to be grouped to coherent instances.

This head type has the advantage that it only needs to evaluate each frame once to find all peaks, in contrast to topdown models that must be evaluated for each crop. This constant scaling with the number of instances can be especially beneficial when there are many animals present in the frame.

edges

List of 2-tuples of the form (source_node, destination_node) that define pairs of text names of the directed edges of the graph. If not set, all edges in the skeleton will be used.

Type

Optional[Sequence[Tuple[str, str]]]

sigma

Spread of the Gaussian distribution that weigh the part affinity fields as a function of their distance from the edge they represent. Smaller values are more precise but may be difficult to learn as they have a lower density within the image space. Larger values are easier to learn but are less precise with respect to the edge distance, so can be less useful in disambiguating between edges that are nearby and parallel in direction. This spread is in units of pixels of the model input image, i.e., the image resolution after any input scaling is applied.

Type

float

output_stride

The stride of the output part affinity fields relative to the input image. This is the reciprocal of the resolution, e.g., an output stride of 2 results in PAFs that are 0.5x the size of the input. Increasing this value can considerably speed up model performance and decrease memory requirements, at the cost of decreased spatial resolution.

Type

int

loss_weight

Scalar float used to weigh the loss term for this head during training. Increase this to encourage the optimization to focus on improving this specific output in multi-head models.

Type

float

class sleap.nn.config.model.PretrainedEncoderConfig(encoder: str = 'efficientnetb0', pretrained: bool = True, decoder_filters: int = 256, decoder_filters_rate: float = 1.0, output_stride: int = 2)[source]

Configuration for UNet backbone with pretrained encoder.

encoder

Name of the network architecture to use as the encoder.

Type

str

pretrained

If True, use initialized with weights pretrained on ImageNet.

Type

bool

decoder_filters

Base number of filters for the upsampling blocks in the decoder.

Type

int

decoder_filters_rate

Factor to scale the number of filters by at each consecutive upsampling block in the decoder.

Type

float

output_stride

Stride of the final output.

Type

int

class sleap.nn.config.model.ResNetConfig(version: str = 'ResNet50', weights: str = 'frozen', upsampling: Optional[sleap.nn.config.model.UpsamplingConfig] = None, max_stride: int = 32, output_stride: int = 4)[source]

ResNet backbone configuration.

version

Name of the ResNetV1 variant. Can be one of: “ResNet50”, “ResNet101”, or “ResNet152”.

Type

str

weights

Controls how the network weights are initialized. If “random”, the network is not pretrained. If “frozen”, the network uses pretrained weights and keeps them fixed. If “tunable”, the network uses pretrained weights and allows them to be trainable.

Type

str

upsampling

A UpsamplingConfig that defines an upsampling branch if not None.

Type

Optional[sleap.nn.config.model.UpsamplingConfig]

max_stride

Stride of the backbone feature activations. These should be <= 32.

Type

int

output_stride

Stride of the final output. If the upsampling branch is not defined, the output stride is controlled via dilated convolutions or reduced pooling in the backbone.

Type

int

class sleap.nn.config.model.SingleInstanceConfmapsHeadConfig(part_names: Optional[List[str]] = None, sigma: float = 5.0, output_stride: int = 1, offset_refinement: bool = False)[source]

Configurations for single instance confidence map heads.

These heads are used in single instance models that make the assumption that only one of each body part is present in the image. These heads produce confidence maps with a single peak for each part type which can be detected via global peak finding.

Do not use this head if there is more than one animal present in the image.

part_names

Text name of the body parts (nodes) that the head will be configured to produce. The number of parts determines the number of channels in the output. If not specified, all body parts in the skeleton will be used.

Type

Optional[List[str]]

sigma

Spread of the Gaussian distribution of the confidence maps as a scalar float. Smaller values are more precise but may be difficult to learn as they have a lower density within the image space. Larger values are easier to learn but are less precise with respect to the peak coordinate. This spread is in units of pixels of the model input image, i.e., the image resolution after any input scaling is applied.

Type

float

output_stride

The stride of the output confidence maps relative to the input image. This is the reciprocal of the resolution, e.g., an output stride of 2 results in confidence maps that are 0.5x the size of the input. Increasing this value can considerably speed up model performance and decrease memory requirements, at the cost of decreased spatial resolution.

Type

int

offset_refinement

If True, model will also output an offset refinement map used to achieve subpixel localization of peaks during inference. This can improve the localization accuracy of the model at the cost of additional memory and training and inference time. If False (the default), subpixel localization can be achieved post-hoc with deterministic refinement, which does not require additional resources or training, but may not achieve the same accuracy as learned refinement.

Type

bool

class sleap.nn.config.model.UNetConfig(stem_stride: Optional[int] = None, max_stride: int = 16, output_stride: int = 1, filters: int = 64, filters_rate: float = 2, middle_block: bool = True, up_interpolate: bool = False, stacks: int = 1)[source]

UNet backbone configuration.

stem_stride

If not None, controls how many stem blocks to use for initial downsampling. These are useful for learned downsampling that is able to retain spatial information while reducing large input image sizes.

Type

Optional[int]

max_stride

Determines the number of downsampling blocks in the network, increasing receptive field size at the cost of network size.

Type

int

output_stride

Determines the number of upsampling blocks in the network.

Type

int

filters

Base number of filters in the network.

Type

int

filters_rate

Factor to scale the number of filters by at each block.

Type

float

middle_block

If True, add an intermediate block between the downsampling and upsampling branch for additional processing for features at the largest receptive field size. This will not introduce an extra pooling step.

Type

bool

up_interpolate

If True, use bilinear upsampling instead of transposed convolutions for upsampling. This can save computations but may lower overall accuracy.

Type

bool

stacks

Number of repeated stacks of the network (excluding the stem).

Type

int

class sleap.nn.config.model.UpsamplingConfig(method: str = 'interpolation', skip_connections: Optional[str] = None, block_stride: int = 2, filters: int = 64, filters_rate: float = 1, refine_convs: int = 2, batch_norm: bool = True, transposed_conv_kernel_size: int = 4)[source]

Upsampling stack configuration.

method

If “transposed_conv”, use a strided transposed convolution to perform learnable upsampling. If “interpolation”, bilinear upsampling will be used instead.

Type

str

skip_connections

If “add”, incoming feature tensors form skip connection with upsampled features via element-wise addition. Height/width are matched via stride and a 1x1 linear conv is applied if the channel counts do no match up. If “concatenate”, the skip connection is formed via channel-wise concatenation. If None, skip connections will not be formed.

Type

Optional[str]

block_stride

The striding of the upsampling layer (not tensor). This is typically set to 2, such that the tensor doubles in size with each upsampling step, but can be set higher to upsample to the desired output_stride directly in fewer steps.

Type

int

filters

Integer that specifies the base number of filters in each convolution layer. This will be scaled by the filters_rate at every upsampling step.

Type

int

filters_rate

Factor to scale the number of filters in the convolution layers after each upsampling step. If set to 1, the number of filters won’t change.

Type

float

refine_convs

If greater than 0, specifies the number of 3x3 convolutions that will be applied after the upsampling step for refinement. These layers can serve the purpose of “mixing” the skip connection fused features, or to refine the current feature map after upsampling, which can help to prevent aliasing and checkerboard effects. If 0, no additional convolutions will be applied.

Type

int

conv_batchnorm

Specifies whether batch norm should be applied after each convolution (and before the ReLU activation).

transposed_conv_kernel_size

Size of the kernel for the transposed convolution. No effect if bilinear upsampling is used.

Type

int