x-transformers 2.1.7__tar.gz → 2.1.10__tar.gz

{x_transformers-2.1.7 → x_transformers-2.1.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.1.7
+Version: 2.1.10
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.1.7 → x_transformers-2.1.10}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.1.7"
+version = "2.1.10"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.1.7 → x_transformers-2.1.10}/tests/test_x_transformers.py

@@ -694,7 +694,10 @@ def test_lime(
 
     model(x)
 
-def test_belief_state_wrapper():
+@pytest.mark.parametrize('backward_ar_loss_weight', (1., 0.5))
+def test_belief_state_wrapper(
+    backward_ar_loss_weight
+):
     from x_transformers.belief_state_wrapper import BeliefStateWrapper
 
     forward_model = TransformerWrapper(
@@ -722,7 +725,7 @@ def test_belief_state_wrapper():
     model = BeliefStateWrapper(
         forward_decoder = forward_model,
         backward_decoder = backward_model,
-        backward_ar_loss_weight = 0.5
+        backward_ar_loss_weight = backward_ar_loss_weight
     )
 
     seq = torch.randint(0, 20000, (2, 16))

{x_transformers-2.1.7 → x_transformers-2.1.10}/x_transformers/belief_state_wrapper.py

@@ -1,8 +1,11 @@
+
 # Belief State Transformer
 
 # https://arxiv.org/abs/2410.23506
 # https://www.youtube.com/watch?v=aqhbRtB2Fyg
 
+from __future__ import annotations
+
 import torch
 from torch.autograd import Function
 from torch.nn import Module, ModuleList
@@ -35,11 +38,14 @@ class BeliefStateWrapper(Module):
     def __init__(
         self,
         forward_decoder: TransformerWrapper,
-        backward_decoder: TransformerWrapper,
+        backward_decoder: TransformerWrapper | None = None,
         train_frac_forward_backward_pairs: float = 1.,
+        text_head: Module | None = None,
         backward_ar_loss_weight: float = 1. # can weigh the training of the backwards decoder differently, perhaps fwd/bwd have a shared backbone etc etc
     ):
         super().__init__()
+        backward_decoder = default(backward_decoder, forward_decoder) # if backward decoder not set, use the same transformer, assume it knows how to switch gears based on suffix token
+
         assert forward_decoder.emb_dim == backward_decoder.emb_dim, 'forward and backwards model must have the same embedding dimension'
         assert forward_decoder.num_tokens == backward_decoder.num_tokens, 'forward and backwards model must have the same number of tokens'
 
@@ -53,11 +59,14 @@ class BeliefStateWrapper(Module):
 
         # the text prediction head, which predicts for the combinations of prefix and suffix the next and previous token for forwards and backward sequences
 
-        self.text_head = nn.Sequential(
-            nn.Linear(dim * 2, dim),
-            nn.LeakyReLU(),
-            nn.Linear(dim, num_tokens * 2),
-        )
+        if not exists(text_head):
+            text_head = nn.Sequential(
+                nn.Linear(dim * 2, dim),
+                nn.LeakyReLU(),
+                nn.Linear(dim, num_tokens * 2),
+            )
+
+        self.text_head = text_head
 
         # the two decoders, one which is causal forward, the other causal backwards