Include MoG and ConditionalNorms.

src/embedding_scvi/_components.py

@@ -8,6 +8,38 @@
 from torch import nn
 
 
+class ConditionalBatchNorm2d(nn.Module):
+    def __init__(self, num_features, num_classes, momentum, eps):
+        super().__init__()
+        self.num_features = num_features
+        self.bn = nn.BatchNorm1d(self.num_features, momentum=momentum, eps=eps, affine=False)
+        self.embed = nn.Embedding(num_classes, self.num_features * 2)
+        self.embed.weight.data[:, :self.num_features].normal_(1, 0.02)  # Initialise scale at N(1, 0.02)
+        self.embed.weight.data[:, self.num_features:].zero_()  # Initialise bias at 0
+
+    def forward(self, x, y):
+        out = self.bn(x)
+        gamma, beta = self.embed(y.long().ravel()).chunk(2, 1)
+        out = gamma.view(-1, self.num_features) * out + beta.view(-1, self.num_features)
+
+        return out
+
+class ConditionalLayerNorm(nn.Module):
+    def __init__(self, num_features, num_classes):
+        super().__init__()
+        self.num_features = num_features
+        self.ln = nn.LayerNorm(self.num_features, elementwise_affine=False)
+        self.embed = nn.Embedding(num_classes, self.num_features * 2)
+        self.embed.weight.data[:, :self.num_features].normal_(1, 0.02)  # Initialise scale at N(1, 0.02)
+        self.embed.weight.data[:, self.num_features:].zero_()  # Initialise bias at 0
+
+    def forward(self, x, y):
+        out = self.ln(x)
+        gamma, beta = self.embed(y.long().ravel()).chunk(2, 1)
+        out = gamma.view(-1, self.num_features) * out + beta.view(-1, self.num_features)
+
+        return out
+
 class MLPBlock(nn.Module):
     """Multi-layer perceptron block.
 
@@ -45,7 +77,9 @@ def __init__(
         self,
         n_in: int,
         n_out: int,
+        cat_dim: int | None = None,
         bias: bool = True,
+        conditional: bool = False,
         norm: Literal["batch", "layer"] | None = None,
         norm_kwargs: dict | None = None,
         activation: Literal["relu", "leaky_relu", "softmax", "softplus"] | None = None,
@@ -62,16 +96,28 @@ def __init__(
         self.dropout = nn.Identity()
         self.residual = residual
 
-        if norm == "batch":
-            self.norm = nn.BatchNorm1d(n_out, **self.norm_kwargs)
-        elif norm == "layer":
-            self.norm = nn.LayerNorm(n_out, **self.norm_kwargs)
-        elif norm is not None:
-            raise InvalidParameterError(
-                param="norm",
-                value=norm,
-                valid=["batch", "layer", None],
-            )
+        if conditional:
+            if norm == "batch":
+                self.norm = ConditionalBatchNorm2d(n_out, cat_dim, momentum=0.01, eps=0.001)
+            elif norm == "layer":
+                self.norm = ConditionalLayerNorm(n_out, cat_dim)
+            elif norm is not None:
+                raise InvalidParameterError(
+                    param="norm",
+                    value=norm,
+                    valid=["batch", "layer", None],
+                )
+        else:
+            if norm == "batch":
+                self.norm = nn.BatchNorm1d(n_out, **self.norm_kwargs)
+            elif norm == "layer":
+                self.norm = nn.LayerNorm(n_out, **self.norm_kwargs)
+            elif norm is not None:
+                raise InvalidParameterError(
+                    param="norm",
+                    value=norm,
+                    valid=["batch", "layer", None],
+                )
 
         if activation == "relu":
             self.activation = nn.ReLU(**self.activation_kwargs)

src/embedding_scvi/_model.py

@@ -38,6 +38,7 @@ def __init__(self, adata: AnnData, **kwargs):
         )
         self.module = EmbeddingVAE(
             n_vars=self.summary_stats.n_vars,
+            n_labels=self.summary_stats.n_labels,
             categorical_covariates=categorical_covariates,
             **kwargs,
         )
@@ -63,6 +64,7 @@ def setup_anndata(
         cls,
         adata: AnnData,
         layer: str | None = None,
+        labels_key: str | None = None,
         categorical_covariate_keys: list[str] | None = None,
         **kwargs,
     ):
@@ -77,6 +79,7 @@ def setup_anndata(
         setup_method_args = cls._get_setup_method_args(**locals())
         anndata_fields = [
             fields.LayerField(REGISTRY_KEYS.X_KEY, layer, is_count_data=True),
+            fields.CategoricalObsField(REGISTRY_KEYS.LABELS_KEY, labels_key),
             ExtendableCategoricalJointObsField(REGISTRY_KEYS.CAT_COVS_KEY, categorical_covariate_keys),
         ]
         adata_manager = AnnDataManager(fields=anndata_fields, setup_method_args=setup_method_args)

src/embedding_scvi/_module.py

@@ -39,10 +39,13 @@ def __init__(
         self,
         n_vars: int,
         n_latent: int = 25,
+        n_labels: int | None = None,
         categorical_covariates: list[int] | None = None,
         likelihood: str = "zinb",
         encoder_kwargs: dict | None = None,
         decoder_kwargs: dict | None = None,
+        prior: str | None = None,
+        mixture_components: int = 50,
     ):
         super().__init__()
 
@@ -53,6 +56,30 @@ def __init__(
         self.encoder_kwargs = encoder_kwargs or {}
         self.decoder_kwargs = decoder_kwargs or {}
 
+        self.prior = prior
+        if self.prior=='mog':
+            self.register_buffer(
+                "u_prior_logits", torch.ones([mixture_components]))
+            self.register_buffer(
+                "u_prior_means", torch.randn([n_latent, mixture_components]))
+            self.register_buffer(
+                "u_prior_scales", torch.zeros([n_latent, mixture_components]))
+        elif self.prior=='mog_celltype':
+            self.register_buffer(
+                "u_prior_logits", torch.ones([n_labels]))
+            self.register_buffer(
+                "u_prior_means", torch.randn([n_latent, n_labels]))
+            self.register_buffer(
+                "u_prior_scales", torch.zeros([n_latent, n_labels]))
+
+        self.covariates_encoder = nn.Identity()
+        if self.categorical_covariates is not None:
+            self.covariates_encoder = ExtendableEmbeddingList(
+                num_embeddings=self.categorical_covariates,
+                embedding_dim=self.n_latent,
+            )
+
+
         encoder_dist_params = likelihood_to_dist_params("normal")
         _encoder_kwargs = {
             "n_hidden": 256,
@@ -62,6 +89,7 @@ def __init__(
             "activation": "gelu",
             "dropout_rate": 0.1,
             "residual": True,
+            "cat_dim": self.categorical_covariates.num_embeddings,
         }
         _encoder_kwargs.update(self.encoder_kwargs)
         self.encoder = MultiOutputMLP(
@@ -81,7 +109,9 @@ def __init__(
             "activation": "gelu",
             "dropout_rate": None,
             "residual": True,
+            "n_out_params": self.categorical_covariates.num_embeddings,
         }
+
         _decoder_kwargs.update(self.decoder_kwargs)
         self.decoder = MultiOutputMLP(
             n_in=self.n_latent,
@@ -91,13 +121,6 @@ def __init__(
             **_decoder_kwargs,
         )
 
-        self.covariates_encoder = nn.Identity()
-        if self.categorical_covariates is not None:
-            self.covariates_encoder = ExtendableEmbeddingList(
-                num_embeddings=self.categorical_covariates,
-                embedding_dim=self.n_latent,
-            )
-
     def get_covariate_embeddings(
         self,
         covariate_indexes: list[int] | int | None,
@@ -113,16 +136,19 @@ def get_covariate_embeddings(
 
     def _get_inference_input(self, tensors: dict[str, torch.Tensor]) -> dict:
         x = tensors[REGISTRY_KEYS.X_KEY]
+        y = tensors[REGISTRY_KEYS.LABELS_KEY]
         covariates = tensors.get(REGISTRY_KEYS.CAT_COVS_KEY, None)
         return {
             REGISTRY_KEYS.X_KEY: x,
+            REGISTRY_KEYS.LABELS_KEY: y,
             REGISTRY_KEYS.CAT_COVS_KEY: covariates,
         }
 
     @auto_move_data
     def inference(
         self,
         X: torch.Tensor,
+        y: torch.Tensor | None = None,
         extra_categorical_covs: torch.Tensor | None = None,
         subset_categorical_covs: int | list[int] | None = None,
     ):
@@ -131,7 +157,23 @@ def inference(
 
         posterior_loc, posterior_scale = self.encoder(X)
         posterior = dist.Normal(posterior_loc, posterior_scale + 1e-9)
-        prior = dist.Normal(torch.zeros_like(posterior_loc), torch.ones_like(posterior_scale))
+
+        if self.prior=='mog':
+            cats = dist.Categorical(logits=self.u_prior_logits)
+            normal_dists = dist.Normal(
+                self.u_prior_means,
+                torch.exp(self.u_prior_scales))
+            prior = dist.MixtureSameFamily(cats, normal_dists)
+        elif self.prior=='mog_celltype':
+            label_bias = 10.0 * torch.nn.functional.one_hot(y, self.n_labels) if self.n_labels >= 2 else 0.0
+            cats = dist.Categorical(logits=self.u_prior_logits + label_bias)
+            normal_dists = dist.Normal(
+                self.u_prior_means,
+                torch.exp(self.u_prior_scales))
+            prior = dist.MixtureSameFamily(cats, normal_dists)
+        else:
+            prior = dist.Normal(torch.zeros_like(posterior_loc), torch.ones_like(posterior_scale))
+
         z = posterior.rsample()
 
         covariates_z = self.covariates_encoder(
@@ -216,10 +258,18 @@ def loss(
         X = tensors[REGISTRY_KEYS.X_KEY]
         posterior = inference_outputs[TENSORS_KEYS.QZ_KEY]
         prior = inference_outputs[TENSORS_KEYS.PZ_KEY]
-        likelihood = generative_outputs[TENSORS_KEYS.PX_KEY]
 
-        # (n_obs, n_latent) -> (n_obs,)
-        kl_div = dist.kl_divergence(posterior, prior).sum(dim=-1)
+        if self.prior=='mog' or self.prior=='mog_celltype':
+            u = posterior.rsample(sample_shape=(10,))
+            # (n_obs, n_latent) -> (n_obs,)
+            kl_z = prior.log_prob(u) - posterior.log_prob(u)
+            kl_div = kl_z.sum(-1)
+        else:
+            # (n_obs, n_latent) -> (n_obs,)
+            kl_div = dist.kl_divergence(posterior, prior).sum(dim=-1)
+
+
+        likelihood = generative_outputs[TENSORS_KEYS.PX_KEY]
         weighted_kl_div = kl_weight * kl_div
         # (n_obs, n_vars) -> (n_obs,)
         reconstruction_loss = -likelihood.log_prob(X).sum(dim=-1)