x-transformers 1.37.3__tar.gz → 1.37.4__tar.gz

{x_transformers-1.37.3/x_transformers.egg-info → x_transformers-1.37.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.37.3
+Version: 1.37.4
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.37.3 → x_transformers-1.37.4}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.37.3',
+  version = '1.37.4',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

{x_transformers-1.37.3 → x_transformers-1.37.4}/x_transformers/nonautoregressive_wrapper.py

@@ -309,9 +309,11 @@ class NonAutoregressiveWrapper(nn.Module):
         with context():
             logits = self.net(masked, **kwargs)
 
+        loss_fn = F.cross_entropy if not self.net.is_log_prob else F.nll_loss
+
         # cross entropy loss
 
-        loss = F.cross_entropy(
+        loss = loss_fn(
             logits[mask],
             orig_seq[mask]
         )

{x_transformers-1.37.3 → x_transformers-1.37.4}/x_transformers/xl_autoregressive_wrapper.py

@@ -40,7 +40,7 @@ class XLAutoregressiveWrapper(nn.Module):
         eos_token = None,
         temperature = 1.,
         filter_logits_fn = top_k,
-        filter_thres = 0.9,
+        filter_kwargs: dict = dict(),
         mems = None,
         **kwargs
     ):
@@ -88,7 +88,7 @@ class XLAutoregressiveWrapper(nn.Module):
             mems = cache.mems
 
             logits = logits[:, -1]
-            filtered_logits = filter_logits_fn(logits, thres = filter_thres)
+            filtered_logits = filter_logits_fn(logits, **filter_kwargs)
             probs = F.softmax(filtered_logits / temperature, dim=-1)
 
             sample = torch.multinomial(probs, 1)
@@ -131,7 +131,9 @@ class XLAutoregressiveWrapper(nn.Module):
 
         split_x = x.split(max_seq_len, dim = -1)
         split_labels = labels.split(max_seq_len, dim = -1)
-        loss_weights = tuple(map(lambda t: t.shape[-1] / seq_len, split_x))
+        loss_weights = tuple((t.shape[-1] / seq_len) for t in split_x)
+
+        loss_fn = F.cross_entropy if not self.net.is_log_prob else F.nll_loss
 
         # go through each chunk and derive weighted losses
 
@@ -146,7 +148,7 @@ class XLAutoregressiveWrapper(nn.Module):
                 **kwargs
             )
 
-            loss = F.cross_entropy(
+            loss = loss_fn(
                 rearrange(logits, 'b n c -> b c n'),
                 chunk_labels,
                 ignore_index = ignore_index

{x_transformers-1.37.3 → x_transformers-1.37.4/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.37.3
+Version: 1.37.4
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang