Only train one task per iter, holding grads

2017-05-25 06:47:42 -05:00 · 2017-05-25 06:47:42 -05:00 · 8500d9b1da
parent b27c587800
commit 8500d9b1da
1 changed files with 3 additions and 2 deletions
--- a/spacy/language.py
+++ b/spacy/language.py
@ -222,8 +222,9 @@ class Language(object):
                continue
            tokvecses, bp_tokvecses = tok2vec.model.begin_update(feats, drop=drop)
            d_tokvecses = proc.update((docs, tokvecses), golds,
-                                      drop=drop, sgd=sgd, losses=losses)
-            bp_tokvecses(d_tokvecses, sgd=sgd)
+                                      drop=drop, sgd=get_grads, losses=losses)
+            bp_tokvecses(d_tokvecses, sgd=get_grads)
+            break
        for key, (W, dW) in grads.items():
            sgd(W, dW, key=key)
        # Clear the tensor variable, to free GPU memory.