Use Link.cleargrads instead of Link.zerograds

as zerograds is deprecated by Chainer
chainer · Aug 25, 2019 · 640ea0d · 640ea0d
1 parent cc0f376
commit 640ea0d
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/chainerrl/agents/reinforce.py b/chainerrl/agents/reinforce.py
@@ -140,7 +140,7 @@ def stop_episode_and_train(self, obs, reward, done=False):
 
     def accumulate_grad(self):
         if self.n_backward == 0:
-            self.model.zerograds()
+            self.model.cleargrads()
         # Compute losses
         losses = []
         for r_seq, log_prob_seq, ent_seq in zip(self.reward_sequences,
@@ -168,7 +168,7 @@ def batch_update(self):
         assert len(self.log_prob_sequences) == self.batchsize
         assert len(self.entropy_sequences) == self.batchsize
         # Update the model
-        self.model.zerograds()
+        assert self.n_backward == 0
         self.accumulate_grad()
         self.optimizer.update()
         self.n_backward = 0