Fix not fully defined shape of gradient accumulators on TF <= 1.9

OpenNMT · Nov 30, 2018 · f6641f1 · f6641f1
1 parent f6cd590
commit f6641f1
Show file tree

Hide file tree

Showing 2 changed files with 5 additions and 1 deletion.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -20,6 +20,7 @@ OpenNMT-tf follows [semantic versioning 2.0.0](https://semver.org/). The API cov
 
 * Checkpoint utilities now save a relative path instead of absolute in the generated checkpoint state
 * Fix error on missing configuration fields that should be optional
+* Fix error on gradient accumulation in TensorFlow versions <= 1.9
 
 ## [1.14.1](https://github.com/OpenNMT/OpenNMT-tf/releases/tag/v1.14.1) (2018-11-28)
 

diff --git a/opennmt/utils/optim.py b/opennmt/utils/optim.py
@@ -209,7 +209,10 @@ def delayed_update(optimizer, grads_and_vars, global_step, accum_count=1):
   accum_grads = []
   accum_grads_and_vars = []
   for grad, var in grads_and_vars:
-    accum_grad = tf.Variable(tf.zeros_like(grad), trainable=False, collections=[])
+    accum_grad = tf.Variable(
+        tf.zeros(var.shape, dtype=grad.dtype),
+        trainable=False,
+        collections=[])
     accum_grads.append(accum_grad)
     accum_grads_and_vars.append((accum_grad, var))