handle disk error

examples/char-rnn/README.md

@@ -115,3 +115,5 @@ weston, improving neural grammatical model flow images is allows belief networks
 generating neural networks, there is not the initial particular marked pseudo-cameral rnns
 sophett, pattern wlth designs for faster than the inference in deep learning. in nips (most),
 ```
+
+See [blog of Andrej Karpathy](http://karpathy.github.io/2015/05/21/rnn-effectiveness/) for more interesting stories on this topic.

tensorpack/callbacks/common.py

@@ -49,11 +49,14 @@ class ModelSaver(Callback):
         return var_dict
 
     def _trigger_epoch(self):
+        try:
             self.saver.save(
                 tf.get_default_session(),
                 self.path,
                 global_step=self.global_step,
                 write_meta_graph=not self.meta_graph_written)
+        except Exception:   # disk error sometimes..
+            logger.exception("Exception in ModelSaver.trigger_epoch!")
         if not self.meta_graph_written:
             self.meta_graph_written = True
 

tensorpack/callbacks/summary.py

@@ -69,9 +69,12 @@ class StatHolder(object):
 
     def _write_stat(self):
         tmp_filename = self.filename + '.tmp'
+        try:
             with open(tmp_filename, 'w') as f:
                 json.dump(self.stat_history, f)
             os.rename(tmp_filename, self.filename)
+        except IOError: # disk error sometimes..
+            logger.exception("Exception in StatHolder.finalize()!")
 
 class StatPrinter(Callback):
     """

tensorpack/train/trainer.py

@@ -91,7 +91,10 @@ class EnqueueThread(threading.Thread):
             except Exception:
                 logger.exception("Exception in EnqueueThread:")
             finally:
+                try:
                     self.sess.run(self.close_op)
+                except RuntimeError:    # session already closed
+                    pass
                 self.coord.request_stop()
                 logger.info("Enqueue Thread Exited.")
 

tensorpack/utils/stat.py

@@ -3,59 +3,63 @@
 # Author: Yuxin Wu <ppwwyyxx@gmail.com>
 import numpy as np
 
-__all__ = ['StatCounter', 'Accuracy', 'BinaryStatistics']
+__all__ = ['StatCounter', 'Accuracy', 'BinaryStatistics', 'RatioStatistics']
 
 class StatCounter(object):
     def __init__(self):
         self.reset()
 
     def feed(self, v):
-        self.values.append(v)
+        self._values.append(v)
 
     def reset(self):
-        self.values = []
+        self._values = []
 
     @property
     def count(self):
-        return len(self.values)
+        return len(self._values)
 
     @property
     def average(self):
-        assert len(self.values)
-        return np.mean(self.values)
+        assert len(self._values)
+        return np.mean(self._values)
 
     @property
     def sum(self):
-        assert len(self.values)
-        return np.sum(self.values)
+        assert len(self._values)
+        return np.sum(self._values)
 
     @property
     def max(self):
-        assert len(self.values)
-        return max(self.values)
+        assert len(self._values)
+        return max(self._values)
 
-class Accuracy(object):
+class RatioStatistics(object):
     def __init__(self):
         self.reset()
 
     def reset(self):
-        self.tot = 0
-        self.corr = 0
+        self._tot = 0
+        self._cnt = 0
 
-    def feed(self, corr, tot=1):
-        self.tot += tot
-        self.corr += corr
+    def feed(self, cnt, tot=1):
+        self._tot += tot
+        self._cnt += cnt
 
     @property
-    def accuracy(self):
-        if self.tot == 0:
+    def ratio(self):
+        if self._tot == 0:
             return 0
-        return self.corr * 1.0 / self.tot
+        return self._cnt * 1.0 / self._tot
 
     @property
     def count(self):
-        return self.tot
+        return self._tot
 
+class Accuracy(RatioStatistics):
+    @property
+    def accuracy(self):
+        return self.ratio
 
 class BinaryStatistics(object):
     """