x-transformers 2.1.25__py3-none-any.whl → 2.1.27__py3-none-any.whl

x_transformers/belief_state_wrapper.py

@@ -9,7 +9,7 @@ from __future__ import annotations
 import torch
 from torch.autograd import Function
 from torch.nn import Module, ModuleList
-from torch import nn, cat, stack, tensor, arange, cartesian_prod
+from torch import nn, cat, stack, tensor, Tensor, arange, cartesian_prod
 import torch.nn.functional as F
 
 from x_transformers.autoregressive_wrapper import (
@@ -34,6 +34,25 @@ def exists(v):
 def default(v, d):
     return v if exists(v) else d
 
+# a custom flip that can handle variable lengths across batch
+
+def flip(x, dim = 1, lens = None):
+    if not exists(lens):
+        return x.flip(dim)
+
+    batch, seq_len, device = *x.shape[:2], x.device
+    seq = arange(seq_len, device = device)
+
+    mask = einx.less('j, i -> i j', seq, lens)
+    masked_seq = einx.where('i j, j,', mask, seq, -1)
+
+    flip_indices = masked_seq.argsort(dim = -1, descending = True)
+
+    if x.ndim == 3:
+        flip_indices = repeat(flip_indices, '... -> ... d', d = x.shape[-1])
+
+    return x.gather(dim, flip_indices)
+
 # wrappers
 
 class BeliefStateWrapper(Module):
@@ -230,19 +249,25 @@ class BeliefStateWrapper(Module):
     def forward(
         self,
         seq,
+        lens: Tensor | None = None, # Int['b']
         return_loss_only = False,
         loss_scale = 1.,
         loss_weight_by_fb_indices: callable | None = None
     ):
         batch, seq_len, device = *seq.shape, seq.device
 
+        # handle variable length sequences
+
+        if exists(lens):
+            mask = einx.less('j, i -> i j', arange(seq_len, device = device), lens)
+
         # forward autoregressive
 
         forward_embeds = self.forward_decoder(seq, return_embeddings = True)
 
         # backward autoregressive
 
-        backward_seq = seq.flip(1)
+        backward_seq = flip(seq, lens = lens)
 
         suffix_tokens = repeat(self.suffix_token, 'd -> b 1 d', b = batch)
 
@@ -252,7 +277,7 @@ class BeliefStateWrapper(Module):
             return_embeddings = True
         )
 
-        backward_embeds = backward_embeds.flip(1)
+        backward_embeds = flip(backward_embeds, lens = lens)
 
         # trick to reduce memory on backwards pass
 
@@ -320,7 +345,7 @@ class BeliefStateWrapper(Module):
         if exists(self.to_distance_logits):
             distance_logits = self.to_distance_logits(fb_embeds)
 
-            distance_labels = bi - fi
+            distance_labels = (bi - fi).clamp(max = self.max_pred_distance - 1)
             distance_labels = repeat(distance_labels, 'n -> b n', b = batch)
 
             pred_dist_loss = F.cross_entropy(

{x_transformers-2.1.25.dist-info → x_transformers-2.1.27.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.1.25
+Version: 2.1.27
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.1.25.dist-info → x_transformers-2.1.27.dist-info}/RECORD

@@ -1,7 +1,7 @@
 x_transformers/__init__.py,sha256=NDoiBivau559WQ0FvXG4ssU3Il9aoHmTIUFN_1juz0s,911
 x_transformers/attend.py,sha256=-5BWWhFsp7tvZTdN91Ay5SqOjyj9uOs-122vFvoO6b4,17253
 x_transformers/autoregressive_wrapper.py,sha256=reLCno9Z9pchVU79tBF8OMo21LwSZ67KAeB83jqkyAc,10505
-x_transformers/belief_state_wrapper.py,sha256=xPzhUZYm7qdHYp9fQ73HjwvWmEhve6-cEisvYK5serI,11571
+x_transformers/belief_state_wrapper.py,sha256=p9OZTX-xs8YAiPdDCpZLqCG8VGmF8OJU6Zriy3mTGfo,12399
 x_transformers/continuous.py,sha256=p0sCAiH1na236ygwgL1Yyhu36eZBf9cZvoW1JyP_fFE,7073
 x_transformers/dpo.py,sha256=xt4OuOWhU8pN3OKN2LZAaC2NC8iiEnchqqcrPWVqf0o,3521
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
@@ -10,7 +10,7 @@ x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dC
 x_transformers/x_transformers.py,sha256=fqgtIs6__JpLWMnJa8AY5OW3AJ2GR1B5p-9TsWdiOIU,110425
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-2.1.25.dist-info/METADATA,sha256=h182tY5ffDUrwUr8VZYekkTeMcDlMD_Krw1XP2K0YWU,87875
-x_transformers-2.1.25.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-x_transformers-2.1.25.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-2.1.25.dist-info/RECORD,,
+x_transformers-2.1.27.dist-info/METADATA,sha256=mUUzC0KPtW-RFPxwHyY_xr72ZrfDjPjjhaCyH4erUTw,87875
+x_transformers-2.1.27.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+x_transformers-2.1.27.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-2.1.27.dist-info/RECORD,,