update gym env

scripts/dump-model-params.py

@@ -49,5 +49,7 @@ with tf.Graph().as_default() as G:
                 var_dict[name] = v
             logger.info("Variables to dump:")
             logger.info(", ".join(var_dict.keys()))
-            saver = tf.train.Saver(var_list=var_dict)
+            saver = tf.train.Saver(
+                    var_list=var_dict,
+                    write_version=tf.train.SaverDef.V2)
             saver.save(sess, args.output, write_meta_graph=False)

tensorpack/RL/gymenv.py

@@ -34,6 +34,7 @@ class GymEnv(RLEnvironment):
         if dumpdir:
             mkdir_p(dumpdir)
             self.gymenv.monitor.start(dumpdir)
+        self.use_dir = dumpdir
 
         self.reset_stat()
         self.rwd_counter = StatCounter()
@@ -46,6 +47,7 @@ class GymEnv(RLEnvironment):
         self._ob = self.gymenv.reset()
 
     def finish_episode(self):
+        if self.use_dir is not None:
             self.gymenv.monitor.flush()
         self.stats['score'].append(self.rwd_counter.sum)
 

tensorpack/callbacks/common.py

@@ -33,6 +33,13 @@ class ModelSaver(Callback):
         for key in self.var_collections:
             vars.extend(tf.get_collection(key))
         self.path = os.path.join(logger.LOG_DIR, 'model')
+        try:
+            self.saver = tf.train.Saver(
+                var_list=ModelSaver._get_var_dict(vars),
+                max_to_keep=self.keep_recent,
+                keep_checkpoint_every_n_hours=self.keep_freq,
+                write_version=tf.train.SaverDef.V2)
+        except:
             self.saver = tf.train.Saver(
                 var_list=ModelSaver._get_var_dict(vars),
                 max_to_keep=self.keep_recent,