Skip to content

novae.module.SwavHead

novae.module.SwavHead

Bases: LightningModule

Source code in novae/module/swav.py 
 20
 21
 22
 23
 24
 25
 26
 27
 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
140
141
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
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
class SwavHead(L.LightningModule):
    queue: None | Tensor  # (n_slides, num_prototypes)

    @utils.format_docs
    def __init__(
        self,
        mode: utils.Mode,
        output_size: int,
        num_prototypes: int,
        temperature: float,
    ):
        """SwavHead module, adapted from the paper ["Unsupervised Learning of Visual Features by Contrasting Cluster Assignments"](https://arxiv.org/abs/2006.09882).

        Args:
            {output_size}
            {num_prototypes}
            {temperature}
        """
        super().__init__()
        self.mode = mode
        self.output_size = output_size
        self.num_prototypes = num_prototypes
        self.temperature = temperature

        self._prototypes = nn.Parameter(torch.empty((self.num_prototypes, self.output_size)))
        self._prototypes = nn.init.kaiming_uniform_(self._prototypes, a=math.sqrt(5), mode="fan_out")
        self.normalize_prototypes()
        self.min_prototypes = 0

        self.queue = None

        self.reset_clustering()

    def set_min_prototypes(self, min_prototypes_ratio: float):
        self.min_prototypes = int(self.num_prototypes * min_prototypes_ratio)

    def init_queue(self, slide_ids: list[str]) -> None:
        """Initialize the slide-queue.

        Args:
            slide_ids: A list of slide ids.
        """
        del self.queue

        shape = (len(slide_ids), Nums.QUEUE_SIZE, self.num_prototypes)
        self.register_buffer("queue", torch.full(shape, 1 / self.num_prototypes))

        self.slide_label_encoder = {slide_id: i for i, slide_id in enumerate(slide_ids)}

    @torch.no_grad()
    def normalize_prototypes(self):
        self.prototypes.data = F.normalize(self.prototypes.data, dim=1, p=2)

    def forward(self, z1: Tensor, z2: Tensor, slide_id: str | None) -> tuple[Tensor, Tensor]:
        """Compute the SwAV loss for two batches of neighborhood graph views.

        Args:
            z1: Batch containing graphs representations `(B, output_size)`
            z2: Batch containing graphs representations `(B, output_size)`

        Returns:
            The SwAV loss, and the mean entropy normalized (for monitoring).
        """
        self.normalize_prototypes()

        projections1 = self.projection(z1)  # (B, K)
        projections2 = self.projection(z2)  # (B, K)

        ilocs = self.prototype_ilocs(projections1, slide_id)

        projections1, projections2 = projections1[:, ilocs], projections2[:, ilocs]

        q1 = self.sinkhorn(projections1)  # (B, K) or (B, len(ilocs))
        q2 = self.sinkhorn(projections2)  # (B, K) or (B, len(ilocs))

        loss = -0.5 * (self.cross_entropy_loss(q1, projections2) + self.cross_entropy_loss(q2, projections1))

        return loss, _mean_entropy_normalized(q1)

    def cross_entropy_loss(self, q: Tensor, p: Tensor) -> Tensor:
        return torch.mean(torch.sum(q * F.log_softmax(p / self.temperature, dim=1), dim=1))

    @utils.format_docs
    def projection(self, z: Tensor) -> Tensor:
        """Compute the projection of the (normalized) representations over the prototypes.

        Args:
            {z}

        Returns:
            The projections of size `(B, K)`.
        """
        z_normalized = F.normalize(z, dim=1, p=2)
        return z_normalized @ self.prototypes.T

    @utils.format_docs
    @torch.no_grad()
    def prototype_ilocs(self, projections: Tensor, slide_id: str | None = None) -> Tensor:
        """Get the indices of the prototypes to use for the current slide.

        Args:
            {projections}
            slide_id: ID of the slide, or `None`.

        Returns:
            The indices of the prototypes to use, or an `Ellipsis` if all prototypes.
        """
        if (self.queue is None) or (slide_id is None) or self.mode.zero_shot:
            return ...

        slide_index = self.slide_label_encoder[slide_id]

        self.queue[slide_index, 1:] = self.queue[slide_index, :-1].clone()
        self.queue[slide_index, 0] = projections.topk(3, dim=0).values[-1]  # top3 more robust than max

        weights, thresholds = self.queue_weights()
        slide_weights = weights[slide_index]

        ilocs = torch.where(slide_weights >= thresholds)[0]
        return ilocs if len(ilocs) >= self.min_prototypes else torch.topk(slide_weights, self.min_prototypes).indices

    def queue_weights(self) -> tuple[Tensor, Tensor]:
        """Convert the queue to a matrix of prototype weight per slide.

        Returns:
            A tensor of shape `(n_slides, K)`.
        """
        max_projections = self.queue.max(dim=1).values

        thresholds = max_projections.max(0).values * Nums.QUEUE_WEIGHT_THRESHOLD_RATIO

        return max_projections, thresholds

    @utils.format_docs
    @torch.no_grad()
    def sinkhorn(self, projections: Tensor) -> Tensor:
        """Apply the Sinkhorn-Knopp algorithm to the projections.

        Args:
            {projections}

        Returns:
            The soft codes from the Sinkhorn-Knopp algorithm, with shape `(B, K)`.
        """
        Q = torch.exp(projections / Nums.SWAV_EPSILON)  # (B, K)
        Q /= torch.sum(Q)

        B, K = Q.shape

        for _ in range(Nums.SINKHORN_ITERATIONS):
            Q /= torch.sum(Q, dim=0, keepdim=True)
            Q /= K
            Q /= torch.sum(Q, dim=1, keepdim=True)
            Q /= B

        return Q / Q.sum(dim=1, keepdim=True)  # ensure rows sum to 1 (for cross-entropy loss)

    def set_kmeans_prototypes(self, latent: np.ndarray):
        kmeans = KMeans(n_clusters=self.num_prototypes, random_state=0, n_init="auto")
        X = latent / (Nums.EPS + np.linalg.norm(latent, axis=1)[:, None])

        kmeans_prototypes = kmeans.fit(X).cluster_centers_
        kmeans_prototypes = kmeans_prototypes / (Nums.EPS + np.linalg.norm(kmeans_prototypes, axis=1)[:, None])

        self._kmeans_prototypes = torch.nn.Parameter(torch.tensor(kmeans_prototypes))

    @property
    def prototypes(self) -> nn.Parameter:
        return self._kmeans_prototypes if self.mode.zero_shot else self._prototypes

    @property
    def clustering(self) -> AgglomerativeClustering:
        clustering_attr = self.mode.clustering_attr

        if getattr(self, clustering_attr) is None:
            self.hierarchical_clustering()

        return getattr(self, clustering_attr)

    @property
    def clusters_levels(self) -> np.ndarray:
        clusters_levels_attr = self.mode.clusters_levels_attr

        if getattr(self, clusters_levels_attr) is None:
            self.hierarchical_clustering()

        return getattr(self, clusters_levels_attr)

    def reset_clustering(self) -> None:
        for attr in self.mode.all_clustering_attrs:
            setattr(self, attr, None)

    def set_clustering(self, clustering: None, clusters_levels: None) -> None:
        setattr(self, self.mode.clustering_attr, clustering)
        setattr(self, self.mode.clusters_levels_attr, clusters_levels)

    def hierarchical_clustering(self) -> None:
        """
        Perform hierarchical clustering on the prototypes. Saves the full tree of clusters.
        """
        X = self.prototypes.data.numpy(force=True)  # (K, O)

        _clustering = AgglomerativeClustering(
            n_clusters=None,
            distance_threshold=0,
            compute_full_tree=True,
            metric="cosine",
            linkage="average",
        )
        _clustering.fit(X)

        _clusters_levels = np.zeros((len(X), len(X)), dtype=np.uint16)
        _clusters_levels[0] = np.arange(len(X))

        for i, (a, b) in enumerate(_clustering.children_):
            clusters = _clusters_levels[i]
            _clusters_levels[i + 1] = clusters
            _clusters_levels[i + 1, np.where((clusters == a) | (clusters == b))] = len(X) + i

        self.set_clustering(_clustering, _clusters_levels)

    def map_leaves_domains(self, series: pd.Series, level: int) -> pd.Series:
        """Map leaves to the parent domain from the corresponding level of the hierarchical tree.

        Args:
            series: Leaves classes
            level: Level of the hierarchical clustering tree (or, number of clusters)

        Returns:
            Series of classes.
        """
        return series.map(lambda x: f"D{self.clusters_levels[-level, int(x[1:])]}" if isinstance(x, str) else x)

    def find_level(self, leaves_indices: np.ndarray, n_domains: int):
        sub_clusters_levels = self.clusters_levels[:, leaves_indices]
        for level in range(1, self.num_prototypes):
            _n_domains = len(np.unique(sub_clusters_levels[-level]))
            if _n_domains == n_domains:
                return level
        raise ValueError(f"Could not find a level with {n_domains=}")

__init__(mode, output_size, num_prototypes, temperature)

SwavHead module, adapted from the paper "Unsupervised Learning of Visual Features by Contrasting Cluster Assignments".

Parameters:

Name Type Description Default
output_size int

Size of the representations, i.e. the encoder outputs (O in the article).

 required
num_prototypes int

Number of prototypes (K in the article).

 required
temperature float

Temperature used in the cross-entropy loss.

 required
Source code in novae/module/swav.py 
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
@utils.format_docs
def __init__(
    self,
    mode: utils.Mode,
    output_size: int,
    num_prototypes: int,
    temperature: float,
):
    """SwavHead module, adapted from the paper ["Unsupervised Learning of Visual Features by Contrasting Cluster Assignments"](https://arxiv.org/abs/2006.09882).

    Args:
        {output_size}
        {num_prototypes}
        {temperature}
    """
    super().__init__()
    self.mode = mode
    self.output_size = output_size
    self.num_prototypes = num_prototypes
    self.temperature = temperature

    self._prototypes = nn.Parameter(torch.empty((self.num_prototypes, self.output_size)))
    self._prototypes = nn.init.kaiming_uniform_(self._prototypes, a=math.sqrt(5), mode="fan_out")
    self.normalize_prototypes()
    self.min_prototypes = 0

    self.queue = None

    self.reset_clustering()

forward(z1, z2, slide_id)

Compute the SwAV loss for two batches of neighborhood graph views.

Parameters:

Name Type Description Default
z1 Tensor

Batch containing graphs representations (B, output_size)

 required
z2 Tensor

Batch containing graphs representations (B, output_size)

 required

Returns:

Type Description
tuple[Tensor, Tensor]

The SwAV loss, and the mean entropy normalized (for monitoring).
Source code in novae/module/swav.py 
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
def forward(self, z1: Tensor, z2: Tensor, slide_id: str | None) -> tuple[Tensor, Tensor]:
    """Compute the SwAV loss for two batches of neighborhood graph views.

    Args:
        z1: Batch containing graphs representations `(B, output_size)`
        z2: Batch containing graphs representations `(B, output_size)`

    Returns:
        The SwAV loss, and the mean entropy normalized (for monitoring).
    """
    self.normalize_prototypes()

    projections1 = self.projection(z1)  # (B, K)
    projections2 = self.projection(z2)  # (B, K)

    ilocs = self.prototype_ilocs(projections1, slide_id)

    projections1, projections2 = projections1[:, ilocs], projections2[:, ilocs]

    q1 = self.sinkhorn(projections1)  # (B, K) or (B, len(ilocs))
    q2 = self.sinkhorn(projections2)  # (B, K) or (B, len(ilocs))

    loss = -0.5 * (self.cross_entropy_loss(q1, projections2) + self.cross_entropy_loss(q2, projections1))

    return loss, _mean_entropy_normalized(q1)

hierarchical_clustering()

Perform hierarchical clustering on the prototypes. Saves the full tree of clusters.

Source code in novae/module/swav.py 
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
def hierarchical_clustering(self) -> None:
    """
    Perform hierarchical clustering on the prototypes. Saves the full tree of clusters.
    """
    X = self.prototypes.data.numpy(force=True)  # (K, O)

    _clustering = AgglomerativeClustering(
        n_clusters=None,
        distance_threshold=0,
        compute_full_tree=True,
        metric="cosine",
        linkage="average",
    )
    _clustering.fit(X)

    _clusters_levels = np.zeros((len(X), len(X)), dtype=np.uint16)
    _clusters_levels[0] = np.arange(len(X))

    for i, (a, b) in enumerate(_clustering.children_):
        clusters = _clusters_levels[i]
        _clusters_levels[i + 1] = clusters
        _clusters_levels[i + 1, np.where((clusters == a) | (clusters == b))] = len(X) + i

    self.set_clustering(_clustering, _clusters_levels)

init_queue(slide_ids)

Initialize the slide-queue.

Parameters:

Name Type Description Default
slide_ids list[str]

A list of slide ids.

 required
Source code in novae/module/swav.py 
56
57
58
59
60
61
62
63
64
65
66
67
def init_queue(self, slide_ids: list[str]) -> None:
    """Initialize the slide-queue.

    Args:
        slide_ids: A list of slide ids.
    """
    del self.queue

    shape = (len(slide_ids), Nums.QUEUE_SIZE, self.num_prototypes)
    self.register_buffer("queue", torch.full(shape, 1 / self.num_prototypes))

    self.slide_label_encoder = {slide_id: i for i, slide_id in enumerate(slide_ids)}

map_leaves_domains(series, level)

Map leaves to the parent domain from the corresponding level of the hierarchical tree.

Parameters:

Name Type Description Default
series Series

Leaves classes

 required
level int

Level of the hierarchical clustering tree (or, number of clusters)

 required

Returns:

Type Description
Series

Series of classes.
Source code in novae/module/swav.py 
241
242
243
244
245
246
247
248
249
250
251
def map_leaves_domains(self, series: pd.Series, level: int) -> pd.Series:
    """Map leaves to the parent domain from the corresponding level of the hierarchical tree.

    Args:
        series: Leaves classes
        level: Level of the hierarchical clustering tree (or, number of clusters)

    Returns:
        Series of classes.
    """
    return series.map(lambda x: f"D{self.clusters_levels[-level, int(x[1:])]}" if isinstance(x, str) else x)

projection(z)

Compute the projection of the (normalized) representations over the prototypes.

Parameters:

Name Type Description Default
z Tensor

The representations of one batch, of size (B, O).

 required

Returns:

Type Description
Tensor

The projections of size (B, K).
Source code in novae/module/swav.py 
102
103
104
105
106
107
108
109
110
111
112
113
@utils.format_docs
def projection(self, z: Tensor) -> Tensor:
    """Compute the projection of the (normalized) representations over the prototypes.

    Args:
        {z}

    Returns:
        The projections of size `(B, K)`.
    """
    z_normalized = F.normalize(z, dim=1, p=2)
    return z_normalized @ self.prototypes.T

prototype_ilocs(projections, slide_id=None)

Get the indices of the prototypes to use for the current slide.

Parameters:

Name Type Description Default
projections Tensor

Projections of the (normalized) representations over the prototypes, of size (B, K).

 required
slide_id str | None

ID of the slide, or None.

 None

Returns:

Type Description
Tensor

The indices of the prototypes to use, or an Ellipsis if all prototypes.
Source code in novae/module/swav.py 
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
@utils.format_docs
@torch.no_grad()
def prototype_ilocs(self, projections: Tensor, slide_id: str | None = None) -> Tensor:
    """Get the indices of the prototypes to use for the current slide.

    Args:
        {projections}
        slide_id: ID of the slide, or `None`.

    Returns:
        The indices of the prototypes to use, or an `Ellipsis` if all prototypes.
    """
    if (self.queue is None) or (slide_id is None) or self.mode.zero_shot:
        return ...

    slide_index = self.slide_label_encoder[slide_id]

    self.queue[slide_index, 1:] = self.queue[slide_index, :-1].clone()
    self.queue[slide_index, 0] = projections.topk(3, dim=0).values[-1]  # top3 more robust than max

    weights, thresholds = self.queue_weights()
    slide_weights = weights[slide_index]

    ilocs = torch.where(slide_weights >= thresholds)[0]
    return ilocs if len(ilocs) >= self.min_prototypes else torch.topk(slide_weights, self.min_prototypes).indices

queue_weights()

Convert the queue to a matrix of prototype weight per slide.

Returns:

Type Description
tuple[Tensor, Tensor]

A tensor of shape (n_slides, K).
Source code in novae/module/swav.py 
141
142
143
144
145
146
147
148
149
150
151
def queue_weights(self) -> tuple[Tensor, Tensor]:
    """Convert the queue to a matrix of prototype weight per slide.

    Returns:
        A tensor of shape `(n_slides, K)`.
    """
    max_projections = self.queue.max(dim=1).values

    thresholds = max_projections.max(0).values * Nums.QUEUE_WEIGHT_THRESHOLD_RATIO

    return max_projections, thresholds

sinkhorn(projections)

Apply the Sinkhorn-Knopp algorithm to the projections.

Parameters:

Name Type Description Default
projections Tensor

Projections of the (normalized) representations over the prototypes, of size (B, K).

 required

Returns:

Type Description
Tensor

The soft codes from the Sinkhorn-Knopp algorithm, with shape (B, K).
Source code in novae/module/swav.py 
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
@utils.format_docs
@torch.no_grad()
def sinkhorn(self, projections: Tensor) -> Tensor:
    """Apply the Sinkhorn-Knopp algorithm to the projections.

    Args:
        {projections}

    Returns:
        The soft codes from the Sinkhorn-Knopp algorithm, with shape `(B, K)`.
    """
    Q = torch.exp(projections / Nums.SWAV_EPSILON)  # (B, K)
    Q /= torch.sum(Q)

    B, K = Q.shape

    for _ in range(Nums.SINKHORN_ITERATIONS):
        Q /= torch.sum(Q, dim=0, keepdim=True)
        Q /= K
        Q /= torch.sum(Q, dim=1, keepdim=True)
        Q /= B

    return Q / Q.sum(dim=1, keepdim=True)  # ensure rows sum to 1 (for cross-entropy loss)