Merge pull request #338 from CatalinVoss/patch-3

Fix top k generation for k != 0

examples/run_gpt2.py

@@ -16,11 +16,17 @@ logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(messa
 logger = logging.getLogger(__name__)
 
 def top_k_logits(logits, k):
+    """
+    Masks everything but the k top entries as -infinity (1e10).
+    Used to mask logits such that e^-infinity -> 0 won't contribute to the
+    sum of the denominator.
+    """
     if k == 0:
         return logits
-    values, _ = torch.topk(logits, k)
-    min_values = values[:, -1]
-    return torch.where(logits < min_values, torch.ones_like(logits, dtype=logits.dtype) * -1e10, logits)
+    else:
+        values = torch.topk(logits, k)[0]
+        batch_mins = values[:, -1].view(-1, 1).expand_as(logits)
+        return torch.where(logits < batch_mins, torch.ones_like(logits) * -1e10, logits)
 
 def sample_sequence(model, length, start_token=None, batch_size=None, context=None, temperature=1, top_k=0, device='cuda', sample=True):
     if start_token is None: