x-transformers 1.26.2__tar.gz → 1.26.4__tar.gz

{x-transformers-1.26.2/x_transformers.egg-info → x-transformers-1.26.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.26.2
+Version: 1.26.4
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x-transformers-1.26.2 → x-transformers-1.26.4}/README.md

@@ -941,6 +941,8 @@ More of Eric's experimental results can be found <a href="https://github.com/bob
 
 You can use this type of relative position if you wish to train at smaller sequence lengths and have it generalize to longer ones, for both autoregressive and bidirectional models.
 
+Update: <a href="https://www.kaggle.com/competitions/stanford-ribonanza-rna-folding/discussion/460121">First place RNA folding using dynamic positional bias</a>
+
 ```python
 import torch
 from x_transformers import TransformerWrapper, Decoder

{x-transformers-1.26.2 → x-transformers-1.26.4}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.26.2',
+  version = '1.26.4',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

{x-transformers-1.26.2 → x-transformers-1.26.4}/x_transformers/autoregressive_wrapper.py

@@ -145,7 +145,7 @@ class AutoregressiveWrapper(Module):
         cache_kv = True,
         **kwargs
     ):
-        max_seq_len, device = self.max_seq_len, prompts.device
+        max_seq_len, greedy, device = self.max_seq_len, temperature == 0., prompts.device
 
         prompts, ps = pack([prompts], '* n')
 
@@ -230,23 +230,30 @@ class AutoregressiveWrapper(Module):
 
             # filter by top_k, top_p (nucleus), top_a, or custom
 
-            filtered_logits = filter_logits_fn(logits, **filter_kwargs)
+            if greedy:
+                sample = logits.argmax(dim = -1, keepdim = True)
+            else:
+                filtered_logits = filter_logits_fn(logits, **filter_kwargs)
+                probs = F.softmax(filtered_logits / temperature, dim=-1)
+                sample = torch.multinomial(probs, 1)
 
-            probs = F.softmax(filtered_logits / temperature, dim=-1)
-
-            sample = torch.multinomial(probs, 1)
+            # concat sample
 
             out = torch.cat((out, sample), dim=-1)
 
-            if exists(eos_token):
-                is_eos_tokens = (out == eos_token)
+            if not exists(eos_token):
+                continue
+
+            is_eos_tokens = (out == eos_token)
+
+            if is_eos_tokens.any(dim = -1).all():
+                break
 
-                if is_eos_tokens.any(dim = -1).all():
-                    # mask out everything after the eos tokens
-                    shifted_is_eos_tokens = F.pad(is_eos_tokens, (1, -1))
-                    mask = shifted_is_eos_tokens.float().cumsum(dim = -1) >= 1
-                    out = out.masked_fill(mask, self.pad_value)
-                    break
+        if exists(eos_token):
+            # mask out everything after the eos tokens
+            shifted_is_eos_tokens = F.pad(is_eos_tokens, (1, -1))
+            mask = shifted_is_eos_tokens.float().cumsum(dim = -1) >= 1
+            out = out.masked_fill(mask, self.pad_value)
 
         out = out[:, t:]
 

{x-transformers-1.26.2 → x-transformers-1.26.4/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.26.2
+Version: 1.26.4
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang