Customizable Embedder and Logit Mapper (lucidrains#288)

* Allow kwargs in TokenEmbedding's forward method

* Move weight init to embedder class

* Allow custom logits mappers

* Add unit tests for embedder and logits mapper

* Revert "Allow custom logits mappers"

This reverts commit e8166b4.

* Use kwargs for token_emb and to_logits

* Revert to version from main

* Add tests for custom token_emb and to_logits

* Undo accidental deletion of **kwargs

* Simplify test_to_logits

* Remove unused import

tests/test_x_transformers.py

@@ -6,7 +6,7 @@
     TransformerWrapper,
     Encoder,
     Decoder,
-    AutoregressiveWrapper,
+    LinearNoBias,
 )
 
 from x_transformers.neo_mlp import (
@@ -394,3 +394,83 @@ def test_custom_alibi():
     pos = torch.tensor([[0, 1, 2, 4], [1, 3, 5, 7]])
 
     logits = model(x, pos = pos)
+
+
+@pytest.mark.parametrize('embedder_type', ('embedding', 'none', 'custom'))
+def test_embedder(embedder_type):
+    num_tokens = 20000
+    dim = 128
+    token_emb_kwargs = {}
+    if embedder_type == 'embedding':
+        embedder = torch.nn.Embedding(num_tokens, dim)
+    elif embedder_type == 'none':
+        embedder = None
+    else:
+        class CustomEmbedder(torch.nn.Module):
+            """
+            Made up embedder that sums two embeddings. Just to check if we can pass additional input to the embedder's
+            forward pass without breaking the model.
+            """
+            def __init__(self, num_tokens, dim):
+                super().__init__()
+                self.embed_x = torch.nn.Embedding(num_tokens, dim)
+                self.embed_y = torch.nn.Embedding(num_tokens, dim)
+            def forward(self, x, y):
+                return self.embed_x(x) + self.embed_y(y)
+            def init_(self):
+                pass
+        embedder = CustomEmbedder(num_tokens, dim)
+        token_emb_kwargs['y'] = torch.randint(0, num_tokens, (2, 1024))
+    model = TransformerWrapper(
+        num_tokens = num_tokens,
+        max_seq_len = 1024,
+        attn_layers = Decoder(
+            dim = dim,
+            depth = 6,
+            heads = 8,
+        ),
+        token_emb = embedder,
+    )
+
+    x = torch.randint(0, 20000, (2, 1024))
+
+    output = model(x, token_emb_kwargs=token_emb_kwargs)
+    assert output.shape == (2, 1024, 20000)
+
+
+@pytest.mark.parametrize("to_logits", ('linear', 'none', 'pointer'))
+def test_to_logits(to_logits):
+    num_tokens = 20000
+    dim = 128
+    to_logits_kwargs = {}
+    if to_logits == 'linear':
+        logit_mapper = LinearNoBias(dim, num_tokens)
+    elif to_logits == 'none':
+        logit_mapper = None
+    else:
+        class PointerNetworkLogits(torch.nn.Module):
+            def __init__(self, dim):
+                super().__init__()
+                self.proj_to_pointers = torch.nn.Linear(dim, dim)
+            def forward(self, model_embeddings, input_embeddings):
+                pointers = self.proj_to_pointers(model_embeddings)
+                logits = torch.matmul(pointers, input_embeddings.permute(0, 2, 1))
+                return logits
+
+        logit_mapper = PointerNetworkLogits(dim)
+        to_logits_kwargs['input_embeddings'] = torch.randn(2, 20000, dim)
+
+    model = TransformerWrapper(
+        num_tokens = num_tokens,
+        max_seq_len = 1024,
+        attn_layers = Decoder(
+            dim = dim,
+            depth = 6,
+            heads = 8,
+        ),
+        to_logits = logit_mapper,
+    )
+
+    x = torch.randint(0, num_tokens, (2, 1024))
+    output = model(x, to_logits_kwargs=to_logits_kwargs)
+    assert output.shape == (2, 1024, 20000)

x_transformers/x_transformers.py

@@ -238,6 +238,13 @@ def forward(self, x):
         token_emb = self.emb(x.long())
         return l2norm(token_emb) if self.l2norm_embed else token_emb
 
+    def init_(self):
+        if self.l2norm_embed:
+            nn.init.normal_(self.emb.weight, std=1e-5)
+            return
+        nn.init.kaiming_normal_(self.emb.weight)
+
+
 # positional embeddings
 
 class AbsolutePositionalEmbedding(Module):
@@ -2261,7 +2268,8 @@ def __init__(
         token_emb: TokenEmbedding | None = None,
         mixture_of_softmax = False,
         mixture_of_softmax_k = 4,
-        sigsoftmax_logits = False
+        sigsoftmax_logits = False,
+        to_logits: Module | None = None,
     ):
         super().__init__()
 
@@ -2363,11 +2371,12 @@ def __init__(
         if return_only_embed:
             self.to_logits = None
         elif tie_embedding:
+            assert isinstance(self.token_emb, TokenEmbedding), 'can only tie embedding if using `TokenEmbedding`'
             self.to_logits = lambda t: t @ self.token_emb.emb.weight.t()
         elif num_output_heads > 1:
             self.to_logits = ModuleList([LinearNoBias(dim, logits_dim) for _ in range(num_output_heads)])
         else:
-            self.to_logits = LinearNoBias(dim, logits_dim)
+            self.to_logits = LinearNoBias(dim, logits_dim) if to_logits is None else to_logits
 
         # memory tokens (like [cls]) from Memory Transformers paper
 
@@ -2388,13 +2397,11 @@ def __init__(
         self.can_cache_kv_outside_max_seq_len = no_abs_pos_emb
 
     def init_(self):
+        if hasattr(self.token_emb, 'init_'):
+            self.token_emb.init_()
         if self.l2norm_embed:
-            nn.init.normal_(self.token_emb.emb.weight, std = 1e-5)
             if not isinstance(self.pos_emb, always):
                 nn.init.normal_(self.pos_emb.emb.weight, std = 1e-5)
-            return
-
-        nn.init.kaiming_normal_(self.token_emb.emb.weight)
 
     def forward(
         self,
@@ -2417,7 +2424,9 @@ def forward(
         attn_z_loss_weight = 1e-4,
         seq_start_pos = None,
         cache: LayerIntermediates | None = None,
-        **kwargs
+        token_emb_kwargs = dict(),
+        to_logits_kwargs = dict(),
+        **kwargs,
     ):
         b, n, device, num_mems, has_memory_tokens, emb_frac_gradient, orig_mask = x.shape[0], x.shape[1], x.device, self.num_memory_tokens, self.num_memory_tokens > 0, self.emb_frac_gradient, mask
 
@@ -2428,7 +2437,7 @@ def forward(
 
         external_pos_emb = exists(pos) and pos.dtype != torch.long
         pos_emb = self.pos_emb(x, pos = pos, seq_start_pos = seq_start_pos) if not external_pos_emb else pos
-        x = self.token_emb(x) + pos_emb
+        x = self.token_emb(x, **token_emb_kwargs) + pos_emb
 
         # add additional embeddings
 
@@ -2583,9 +2592,9 @@ def forward(
 
         if not return_embeddings:
             if self.has_multiple_heads:
-                logits = tuple(fn(x) for fn in self.to_logits)
+                logits = tuple(fn(x, **to_logits_kwargs) for fn in self.to_logits)
             else:
-                logits = self.to_logits(x)
+                logits = self.to_logits(x, **to_logits_kwargs)
 
         # maybe sig softmax