Use numpy.asarray() instead of numpy.array() when converting feeds.

mrry · tensorflower-gardener · commit f68cbf19fb8f · 2016-06-20T19:33:03.000-07:00
This reduces the number of unnecessary NumPy array allocations when feeding values into a TensorFlow session, and can reduce heap fragmentation. Fixes tensorflow#2942 (as far as possible). Note that we recommend that you use tcmalloc when running TensorFlow, as it is less susceptible to heap fragmentation with mmany large objects. Change: 125407787
diff --git a/tensorflow/python/client/session.py b/tensorflow/python/client/session.py
@@ -616,7 +616,7 @@ def _feed_fn(feed, feed_val):
                 ' Try explicitly setting the type of the feed tensor'
                 ' to a larger type (e.g. int64).')
 
-          np_val = np.array(subfeed_val, dtype=subfeed_dtype)
+          np_val = np.asarray(subfeed_val, dtype=subfeed_dtype)
 
           if not subfeed_t.get_shape().is_compatible_with(np_val.shape):
             raise ValueError(
