x-transformers 2.1.23__py3-none-any.whl → 2.1.25__py3-none-any.whl

x_transformers/belief_state_wrapper.py

@@ -48,8 +48,9 @@ class BeliefStateWrapper(Module):
         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
-        pred_terminal = False,
-        pred_terminal_loss_weight: float = 1.
+        pred_distance = False,
+        pred_distance_loss_weight: float = 1.,
+        max_pred_distance = None
     ):
         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
@@ -59,6 +60,7 @@ class BeliefStateWrapper(Module):
 
         dim = forward_decoder.emb_dim
         num_tokens = forward_decoder.num_tokens
+        max_seq_len = forward_decoder.max_seq_len
 
         self.num_tokens = num_tokens
 
@@ -80,14 +82,15 @@ class BeliefStateWrapper(Module):
 
         # predicting terminal state (when suffix and prefix predict the same token)
 
-        self.to_terminal_logit = nn.Sequential(
+        self.max_pred_distance = default(max_pred_distance, max_seq_len)
+
+        self.to_distance_logits = nn.Sequential(
             nn.Linear(dim * 2, dim),
             nn.LeakyReLU(),
-            nn.Linear(dim, 1),
-            Rearrange('... 1 -> ...')
-        ) if pred_terminal else None
+            nn.Linear(dim, self.max_pred_distance),
+        ) if pred_distance else None
 
-        self.pred_terminal_loss_weight = pred_terminal_loss_weight
+        self.pred_distance_loss_weight = pred_distance_loss_weight
 
         # the two decoders, one which is causal forward, the other causal backwards
 
@@ -314,20 +317,20 @@ class BeliefStateWrapper(Module):
 
         # maybe predict terminal
 
-        if exists(self.to_terminal_logit):
-            terminal_logits = self.to_terminal_logit(fb_embeds)
+        if exists(self.to_distance_logits):
+            distance_logits = self.to_distance_logits(fb_embeds)
 
-            terminal_labels = ((bi - fi) == 2).float() # distance is exactly 2
-            terminal_labels = repeat(terminal_labels, 'n -> b n', b = batch)
+            distance_labels = bi - fi
+            distance_labels = repeat(distance_labels, 'n -> b n', b = batch)
 
-            is_end_loss = F.binary_cross_entropy_with_logits(
-                terminal_logits,
-                terminal_labels
+            pred_dist_loss = F.cross_entropy(
+                rearrange(distance_logits, 'b n l -> b l n'),
+                distance_labels
             )
 
             loss = (
                 loss +
-                is_end_loss * self.pred_terminal_loss_weight
+                pred_dist_loss * self.pred_distance_loss_weight
             )
 
         # maybe early return loss
@@ -350,7 +353,13 @@ class BeliefStateWrapper(Module):
 
             if exists(loss_weight_by_fb_indices):
                 loss_weight = loss_weight_by_fb_indices(fb_pairs)
-                loss = einx.multiply('b fb n, n', loss, loss_weight)
+
+                if loss_weight.ndim == 1:
+                    loss = einx.multiply('b fb n, n', loss, loss_weight)
+                elif loss_weight.ndim == 2:
+                    loss = einx.multiply('b fb n, n fb', loss, loss_weight)
+                else:
+                    raise ValueError('invalid loss weight dims')
 
             loss = loss.mean()
 

{x_transformers-2.1.23.dist-info → x_transformers-2.1.25.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.1.23
+Version: 2.1.25
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.1.23.dist-info → x_transformers-2.1.25.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=qDtASUlmc-JcUM0u7pALFgjS9aEsis-jgL07gxWnjsg,11198
+x_transformers/belief_state_wrapper.py,sha256=xPzhUZYm7qdHYp9fQ73HjwvWmEhve6-cEisvYK5serI,11571
 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-2.1.23.dist-info/METADATA,sha256=26PtwpKeLHzR8yI_4ezvQvmuvz-P3uqaYcPK_RVNvcU,87875
-x_transformers-2.1.23.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-x_transformers-2.1.23.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-2.1.23.dist-info/RECORD,,
+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,,