Added reward logging

examples/A3C-Gym/train-atari.py

@@ -151,6 +151,11 @@ class MySimulatorMaster(SimulatorMaster, Callback):
         super(MySimulatorMaster, self).__init__(pipe_c2s, pipe_s2c)
         self.queue = queue.Queue(maxsize=BATCH_SIZE * 8 * 2)
         self._gpus = gpus
+        self.reward = 0
+        self.fd = open('/kaggle/working', 'w')
+
+    def __del__(self):
+        self.f.close()
 
     def _setup_graph(self):
         # Create predictors on the available predictor GPUs.
@@ -195,8 +200,11 @@ class MySimulatorMaster(SimulatorMaster, Callback):
             client.memory[-1].reward = reward
             if isOver:
                 # should clear client's memory and put to queue
+                self.f.write(str(self.reward) + '\n')
+                self.reward = 0
                 self._parse_memory(0, client, True)
             else:
+                self.reward += reward
                 if len(client.memory) == LOCAL_TIME_MAX + 1:
                     R = client.memory[-1].value
                     self._parse_memory(R, client, False)
@@ -269,7 +277,7 @@ def train():
         session_creator=sesscreate.NewSessionCreator(config=get_default_sess_config(0.5)),
         steps_per_epoch=STEPS_PER_EPOCH,
         session_init=SmartInit(args.load),
-        max_epoch=1000,
+        max_epoch=15,
     )
     trainer = SimpleTrainer() if num_gpu == 1 else AsyncMultiGPUTrainer(train_tower)
     launch_train_with_config(config, trainer)