stats for RLEnv

examples/Atari2600/DQN.py

@@ -255,8 +255,8 @@ def get_config(romfile):
             output = output.strip()
             output = output[output.find(']')+1:]
             mean, maximum = re.findall('[0-9\.]+', output)
-            self.trainer.write_scalar_summary('mean_score', mean)
-            self.trainer.write_scalar_summary('max_score', maximum)
+            self.trainer.write_scalar_summary('eval_mean_score', mean)
+            self.trainer.write_scalar_summary('eval_max_score', maximum)
 
     return TrainConfig(
         dataset=dataset_train,

tensorpack/dataflow/RL.py

@@ -6,9 +6,10 @@
 from abc import abstractmethod, ABCMeta
 import random
 import numpy as np
-from collections import deque, namedtuple
+from collections import deque, namedtuple, defaultdict
 from tqdm import tqdm
 import cv2
+import six
 
 from .base import DataFlow
 from tensorpack.utils import *
@@ -26,6 +27,9 @@ Experience = namedtuple('Experience',
 class RLEnvironment(object):
     __meta__ = ABCMeta
 
+    def __init__(self):
+        self.reset_stat()
+
     @abstractmethod
     def current_state(self):
         """
@@ -40,6 +44,16 @@ class RLEnvironment(object):
         :returns: (reward, isOver)
         """
 
+    @abstractmethod
+    def get_stat(self):
+        """
+        return a dict of statistics (e.g., score) after running for a while
+        """
+
+    def reset_stat(self):
+        """ reset the statistics counter"""
+        self.stats = defaultdict(list)
+
 class NaiveRLEnvironment(RLEnvironment):
     """ for testing only"""
     def __init__(self):
@@ -67,7 +81,9 @@ class ExpReplay(DataFlow, Callback):
             exploration=1,
             end_exploration=0.1,
             exploration_epoch_anneal=0.002,
-            reward_clip=None):
+            reward_clip=None,
+            new_experience_per_step=1
+            ):
         """
         :param predictor: callabale. called with a state, return a distribution
         :param player: a `RLEnvironment`
@@ -117,7 +133,8 @@ class ExpReplay(DataFlow, Callback):
             idxs = self.rng.randint(len(self.mem), size=self.batch_size)
             batch_exp = [self.mem[k] for k in idxs]
             yield self._process_batch(batch_exp)
-            self._populate_exp()
+            for _ in range(self.new_experience_per_step):
+                self._populate_exp()
 
     def _process_batch(self, batch_exp):
         state_shape = batch_exp[0].state.shape
@@ -144,7 +161,11 @@ class ExpReplay(DataFlow, Callback):
         if self.exploration > self.end_exploration:
             self.exploration -= self.exploration_epoch_anneal
             logger.info("Exploration changed to {}".format(self.exploration))
-
+        stats = self.player.get_stat()
+        for k, v in six.iteritems(stats):
+            if isinstance(v, float):
+                self.trainer.write_scalar_summary('expreplay/' + k, v)
+        self.player.reset_stat()
 
 
 if __name__ == '__main__':

tensorpack/dataflow/dataset/atari.py

@@ -107,17 +107,20 @@ class AtariPlayer(RLEnvironment):
         :param action_repeat: repeat each action `action_repeat` times and skip those frames
         :param image_shape: the shape of the observed image
         """
+        super(AtariPlayer, self).__init__()
         for k, v in locals().items():
             if k != 'self':
                 setattr(self, k, v)
         self.last_act = 0
         self.frames = deque(maxlen=hist_len)
+        self.current_accum_score = 0
         self.restart()
 
     def restart(self):
         """
         Restart the game and populate frames with the beginning frame
         """
+        self.current_accum_score = 0
         self.frames.clear()
         s = self.driver.grab_image()
 
@@ -156,11 +159,22 @@ class AtariPlayer(RLEnvironment):
             if isOver:
                 break
         s = cv2.resize(s, self.image_shape)
+        self.current_accum_score += totr
         self.frames.append(s)
         if isOver:
+            self.stats['score'].append(self.current_accum_score)
             self.restart()
         return (totr, isOver)
 
+    def get_stat(self):
+        try:
+            print self.stats
+            return {'avg_score': np.mean(self.stats['score']),
+                    'max_score': float(np.max(self.stats['score']))
+                    }
+        except ValueError:
+            return {}
+
 if __name__ == '__main__':
     a = AtariDriver('breakout.bin', viz=True)
     num = a.get_num_actions()