better eval

examples/Atari2600/DQN.py

@@ -206,7 +206,7 @@ def eval_model_multiprocess(model_path, romfile):
                     self.outq.put(tot_reward)
                     tot_reward = 0
 
-    NR_PROC = multiprocessing.cpu_count() // 2
+    NR_PROC = min(multiprocessing.cpu_count() // 2, 10)
     procs = []
     q = multiprocessing.Queue()
     for k in range(NR_PROC):

tensorpack/predict.py

@@ -116,6 +116,7 @@ class ParallelPredictWorker(multiprocessing.Process):
             os.environ['CUDA_VISIBLE_DEVICES'] = self.gpuid
         else:
             logger.info("Worker {} uses CPU".format(self.idx))
+            os.environ['CUDA_VISIBLE_DEVICES'] = ''
         G = tf.Graph()     # build a graph for each process, because they don't need to share anything
         with G.as_default(), tf.device('/gpu:0' if self.gpuid >= 0 else '/cpu:0'):
             if self.idx != 0: