Skip to content

topdown_multiclass

sleap_nn.inference.layers.topdown_multiclass

TopDownMultiClassLayer — multi-class variant of top-down inference.

Composes :class:CentroidLayer with a multi-class centered-instance CenteredInstanceMultiClassLayer (defined in this module). The multi-class centered-instance model emits keypoint confmaps plus a ClassVectorsHead per crop; the layer adds a class-prob output to Outputs.instance_scores and a per-node class-index field.

Classes:

Name Description
CenteredInstanceMultiClassLayer

Centered-instance + per-instance class vector head.
TopDownMultiClassLayer

Top-down with per-instance class identity.

CenteredInstanceMultiClassLayer

Bases: InferenceLayer

Centered-instance + per-instance class vector head.

Per-crop, returns keypoints (like CenteredInstanceLayer) plus a class index and class probability from a softmax-classified ClassVectorsHead.

Parameters:

Name Type Description Default
backend ModelBackend

Runtime backend wrapping the multi-class centered- instance Lightning module.

 required
output_stride int

Confmap output stride.

 required
max_stride int

Max stride for input divisibility (preprocess pads bottom-right).

 1
preprocess_config / postprocess_config

Standard knobs.

 required
class_names Optional[list[str]]

Ordered class names from multi_class_topdown.class_vectors.classes. Used by the predictor to build the sio.Track registry for identity packaging; an instance's class index (from ClassVectorsHead) indexes into this list.

 None
Notes

Class-level use_gt_peaks = False: the multi-class variant does not support the GT-keypoints fallback path (you can't match GT classes to GT keypoints the way the plain centered-instance layer matches centroids → keypoints). The attribute is exposed so :class:TopDownLayer-derived composers can branch on it uniformly without isinstance checks.

Methods:

Name Description
__init__

Compose the layer with the standard centered-instance config.
postprocess

Decode confmaps to keypoints; classify via ClassVectorsHead.
Source code in sleap_nn/inference/layers/topdown_multiclass.py 
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
class CenteredInstanceMultiClassLayer(InferenceLayer):
    """Centered-instance + per-instance class vector head.

    Per-crop, returns keypoints (like ``CenteredInstanceLayer``) plus a
    class index and class probability from a softmax-classified
    ``ClassVectorsHead``.

    Args:
        backend: Runtime backend wrapping the multi-class centered-
            instance Lightning module.
        output_stride: Confmap output stride.
        max_stride: Max stride for input divisibility (preprocess pads
            bottom-right).
        preprocess_config / postprocess_config: Standard knobs.
        class_names: Ordered class names from
            ``multi_class_topdown.class_vectors.classes``. Used by the
            predictor to build the ``sio.Track`` registry for identity
            packaging; an instance's class index (from ``ClassVectorsHead``)
            indexes into this list.

    Notes:
        Class-level ``use_gt_peaks = False``: the multi-class variant does
        not support the GT-keypoints fallback path (you can't match GT
        classes to GT keypoints the way the plain centered-instance layer
        matches centroids → keypoints). The attribute is exposed so
        :class:`TopDownLayer`-derived composers can branch on it
        uniformly without isinstance checks.
    """

    use_gt_peaks: bool = False

    def __init__(
        self,
        backend: ModelBackend,
        output_stride: int,
        max_stride: int = 1,
        preprocess_config: Optional[PreprocessConfig] = None,
        postprocess_config: Optional[PostprocessConfig] = None,
        class_names: Optional[list[str]] = None,
        class_output: str = "track",
    ) -> None:
        """Compose the layer with the standard centered-instance config."""
        super().__init__(
            backend=backend,
            preprocess_config=preprocess_config or PreprocessConfig(),
            postprocess_config=postprocess_config or PostprocessConfig(),
            output_stride=output_stride,
            max_stride=max_stride,
        )
        self.class_names = list(class_names) if class_names is not None else None
        self.class_output = class_output

    def postprocess(self, raw_out: dict, info: PreprocInfo) -> Outputs:
        """Decode confmaps to keypoints; classify via ``ClassVectorsHead``."""
        cms = raw_out["CenteredInstanceConfmapsHead"]
        peak_class_probs = raw_out["ClassVectorsHead"]  # (n_crops, n_classes)

        peaks, vals = find_global_peaks(
            cms.detach(),
            threshold=self.postprocess_config.peak_threshold,
            refinement=self.postprocess_config.effective_refinement,
            integral_patch_size=self.postprocess_config.integral_patch_size,
        )
        peaks = peaks * info.output_stride
        if info.input_scale != 1.0:
            peaks = peaks / info.input_scale
        eff = info.eff_scale.to(peaks.device)
        if not torch.all(eff == 1.0):
            peaks = peaks / eff.view(-1, 1, 1)

        # NOTE on frame grouping: the legacy multi-class top-down classifies
        # crops PER FRAME — ``get_class_inds_from_vectors`` runs Hungarian
        # assignment over each frame's crops independently (topdown.py:723-733
        # is called once per frame via ``CentroidCrop._generate_crops``, which
        # builds one ``instance_image`` batch per frame). When this layer is
        # composed inside ``TopDownMultiClassLayer``, crops from *all* frames in
        # a batch arrive flattened in a single forward, so classifying them
        # jointly here would mis-assign across frame boundaries. We therefore
        # carry the raw per-crop class vectors in ``pred_class_probs`` and let
        # ``TopDownLayer._run_stage_2`` redo the assignment per frame using the
        # crop→frame mapping it owns. The pre-classified fields below are still
        # populated for the standalone (single-frame) layer path.
        class_inds, class_probs = get_class_inds_from_vectors(peak_class_probs)

        # Reshape peaks to canonical (B=n_crops, I=1, N, 2) Outputs shape.
        peaks_BIN2 = peaks.unsqueeze(1)
        vals_BIN = vals.unsqueeze(1)
        class_inds_BIN = class_inds.unsqueeze(1)  # (n_crops, 1)
        class_probs_BI = class_probs.unsqueeze(1).to(peaks.device)  # (n_crops, 1)
        # Raw per-crop softmax vectors, canonical (n_crops, I=1, n_classes).
        class_vectors_BIC = peak_class_probs.detach().unsqueeze(1).to(peaks.device)

        # Legacy parity (predictors.py:3808-3880): the per-instance ``score``
        # is the centroid confidence (added in ``TopDownLayer._run_stage_2``),
        # and the class probability is the ``tracking_score``. So carry the
        # class prob as ``instance_tracking_scores`` here and leave
        # ``instance_scores`` unset so stage 2 falls back to ``centroid_vals``.
        outputs = Outputs(
            pred_keypoints=peaks_BIN2,
            pred_peak_values=vals_BIN,
            pred_class_inds=class_inds_BIN.unsqueeze(-1).expand(-1, -1, peaks.shape[1]),
            pred_class_probs=class_vectors_BIC,
            instance_tracking_scores=class_probs_BI,
            preprocess_info=info,
        )
        if self.postprocess_config.return_confmaps:
            outputs = attrs.evolve(outputs, pred_confmaps=cms.detach())
        if self.postprocess_config.return_class_vectors:
            outputs = attrs.evolve(
                outputs, pred_class_vectors=peak_class_probs.detach()
            )
        return outputs

__init__(backend, output_stride, max_stride=1, preprocess_config=None, postprocess_config=None, class_names=None, class_output='track')

Compose the layer with the standard centered-instance config.

Source code in sleap_nn/inference/layers/topdown_multiclass.py 
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
def __init__(
    self,
    backend: ModelBackend,
    output_stride: int,
    max_stride: int = 1,
    preprocess_config: Optional[PreprocessConfig] = None,
    postprocess_config: Optional[PostprocessConfig] = None,
    class_names: Optional[list[str]] = None,
    class_output: str = "track",
) -> None:
    """Compose the layer with the standard centered-instance config."""
    super().__init__(
        backend=backend,
        preprocess_config=preprocess_config or PreprocessConfig(),
        postprocess_config=postprocess_config or PostprocessConfig(),
        output_stride=output_stride,
        max_stride=max_stride,
    )
    self.class_names = list(class_names) if class_names is not None else None
    self.class_output = class_output

postprocess(raw_out, info)

Decode confmaps to keypoints; classify via ClassVectorsHead.

Source code in sleap_nn/inference/layers/topdown_multiclass.py 
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
def postprocess(self, raw_out: dict, info: PreprocInfo) -> Outputs:
    """Decode confmaps to keypoints; classify via ``ClassVectorsHead``."""
    cms = raw_out["CenteredInstanceConfmapsHead"]
    peak_class_probs = raw_out["ClassVectorsHead"]  # (n_crops, n_classes)

    peaks, vals = find_global_peaks(
        cms.detach(),
        threshold=self.postprocess_config.peak_threshold,
        refinement=self.postprocess_config.effective_refinement,
        integral_patch_size=self.postprocess_config.integral_patch_size,
    )
    peaks = peaks * info.output_stride
    if info.input_scale != 1.0:
        peaks = peaks / info.input_scale
    eff = info.eff_scale.to(peaks.device)
    if not torch.all(eff == 1.0):
        peaks = peaks / eff.view(-1, 1, 1)

    # NOTE on frame grouping: the legacy multi-class top-down classifies
    # crops PER FRAME — ``get_class_inds_from_vectors`` runs Hungarian
    # assignment over each frame's crops independently (topdown.py:723-733
    # is called once per frame via ``CentroidCrop._generate_crops``, which
    # builds one ``instance_image`` batch per frame). When this layer is
    # composed inside ``TopDownMultiClassLayer``, crops from *all* frames in
    # a batch arrive flattened in a single forward, so classifying them
    # jointly here would mis-assign across frame boundaries. We therefore
    # carry the raw per-crop class vectors in ``pred_class_probs`` and let
    # ``TopDownLayer._run_stage_2`` redo the assignment per frame using the
    # crop→frame mapping it owns. The pre-classified fields below are still
    # populated for the standalone (single-frame) layer path.
    class_inds, class_probs = get_class_inds_from_vectors(peak_class_probs)

    # Reshape peaks to canonical (B=n_crops, I=1, N, 2) Outputs shape.
    peaks_BIN2 = peaks.unsqueeze(1)
    vals_BIN = vals.unsqueeze(1)
    class_inds_BIN = class_inds.unsqueeze(1)  # (n_crops, 1)
    class_probs_BI = class_probs.unsqueeze(1).to(peaks.device)  # (n_crops, 1)
    # Raw per-crop softmax vectors, canonical (n_crops, I=1, n_classes).
    class_vectors_BIC = peak_class_probs.detach().unsqueeze(1).to(peaks.device)

    # Legacy parity (predictors.py:3808-3880): the per-instance ``score``
    # is the centroid confidence (added in ``TopDownLayer._run_stage_2``),
    # and the class probability is the ``tracking_score``. So carry the
    # class prob as ``instance_tracking_scores`` here and leave
    # ``instance_scores`` unset so stage 2 falls back to ``centroid_vals``.
    outputs = Outputs(
        pred_keypoints=peaks_BIN2,
        pred_peak_values=vals_BIN,
        pred_class_inds=class_inds_BIN.unsqueeze(-1).expand(-1, -1, peaks.shape[1]),
        pred_class_probs=class_vectors_BIC,
        instance_tracking_scores=class_probs_BI,
        preprocess_info=info,
    )
    if self.postprocess_config.return_confmaps:
        outputs = attrs.evolve(outputs, pred_confmaps=cms.detach())
    if self.postprocess_config.return_class_vectors:
        outputs = attrs.evolve(
            outputs, pred_class_vectors=peak_class_probs.detach()
        )
    return outputs

TopDownMultiClassLayer

Bases: TopDownLayer

Top-down with per-instance class identity.

Same composition as :class:TopDownLayer (stage 1 = centroid, stage 2 = centered-instance) but the centered-instance layer is a :class:CenteredInstanceMultiClassLayer whose output carries class indices + probabilities. The composition logic is unchanged — class fields propagate through stage 2 unchanged.

Methods:

Name Description
__init__

Forward to TopDownLayer after type-checking the inner layer.

Attributes:

Name Type Description
class_names Optional[list[str]]

Class names from the inner multi-class centered-instance layer.
class_output str

Class-output mode from the inner multi-class centered-instance layer.
Source code in sleap_nn/inference/layers/topdown_multiclass.py 
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
class TopDownMultiClassLayer(TopDownLayer):
    """Top-down with per-instance class identity.

    Same composition as :class:`TopDownLayer` (stage 1 = centroid, stage
    2 = centered-instance) but the centered-instance layer is a
    :class:`CenteredInstanceMultiClassLayer` whose output carries class
    indices + probabilities. The composition logic is unchanged — class
    fields propagate through stage 2 unchanged.
    """

    def __init__(
        self,
        centroid_layer: CentroidLayer,
        centered_instance_layer: CenteredInstanceMultiClassLayer,
        crop_size: Tuple[int, int],
        centroid_nms: bool = False,
        centroid_nms_threshold: float = 0.5,
        return_crops: bool = False,
    ) -> None:
        """Forward to ``TopDownLayer`` after type-checking the inner layer."""
        if not isinstance(centered_instance_layer, CenteredInstanceMultiClassLayer):
            raise TypeError(
                "TopDownMultiClassLayer requires a CenteredInstanceMultiClassLayer "
                "for the centered_instance_layer argument; got "
                f"{type(centered_instance_layer).__name__}."
            )
        super().__init__(
            centroid_layer=centroid_layer,
            centered_instance_layer=centered_instance_layer,
            crop_size=crop_size,
            centroid_nms=centroid_nms,
            centroid_nms_threshold=centroid_nms_threshold,
            return_crops=return_crops,
        )

    @property
    def class_names(self) -> Optional[list[str]]:
        """Class names from the inner multi-class centered-instance layer."""
        return self.centered_instance_layer.class_names

    @property
    def class_output(self) -> str:
        """Class-output mode from the inner multi-class centered-instance layer."""
        return getattr(self.centered_instance_layer, "class_output", "track")

class_names property

Class names from the inner multi-class centered-instance layer.

class_output property

Class-output mode from the inner multi-class centered-instance layer.

__init__(centroid_layer, centered_instance_layer, crop_size, centroid_nms=False, centroid_nms_threshold=0.5, return_crops=False)

Forward to TopDownLayer after type-checking the inner layer.

Source code in sleap_nn/inference/layers/topdown_multiclass.py 
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
def __init__(
    self,
    centroid_layer: CentroidLayer,
    centered_instance_layer: CenteredInstanceMultiClassLayer,
    crop_size: Tuple[int, int],
    centroid_nms: bool = False,
    centroid_nms_threshold: float = 0.5,
    return_crops: bool = False,
) -> None:
    """Forward to ``TopDownLayer`` after type-checking the inner layer."""
    if not isinstance(centered_instance_layer, CenteredInstanceMultiClassLayer):
        raise TypeError(
            "TopDownMultiClassLayer requires a CenteredInstanceMultiClassLayer "
            "for the centered_instance_layer argument; got "
            f"{type(centered_instance_layer).__name__}."
        )
    super().__init__(
        centroid_layer=centroid_layer,
        centered_instance_layer=centered_instance_layer,
        crop_size=crop_size,
        centroid_nms=centroid_nms,
        centroid_nms_threshold=centroid_nms_threshold,
        return_crops=return_crops,
    )