fix PER CUDA device assignment (#171)

* fix PER CUDA device assignment * fix reference * efficient dynamic device setting * set tensor device * mvoe to cpu instead for efficiency
kengz · Sep 15, 2018 · bdb4942 · bdb4942
1 parent 66788f9
commit bdb4942
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 1 deletion.
diff --git a/slm_lab/agent/memory/prioritized.py b/slm_lab/agent/memory/prioritized.py
@@ -138,7 +138,7 @@ def add_experience(self, state, action, reward, next_state, done, error=100000):
 
     def get_priority(self, error):
         '''Takes in the error of one or more examples and returns the proportional priority'''
-        p = torch.pow(error + self.epsilon, self.alpha)
+        p = torch.pow(error.cpu() + self.epsilon, self.alpha)
         return p.squeeze_().cpu().numpy()
 
     def sample_idxs(self, batch_size):

diff --git a/slm_lab/experiment/monitor.py b/slm_lab/experiment/monitor.py
@@ -81,6 +81,7 @@ class Body:
 
     def __init__(self, env, agent_spec, aeb=(0, 0, 0), aeb_space=None):
         # essential reference variables
+        self.agent = None  # set later
         self.env = env
         self.aeb = aeb
         self.a, self.e, self.b = aeb