use video and fix a bug

examples/OpenAIGym/README.md

@@ -5,12 +5,12 @@
 2. Download models from [model zoo](https://drive.google.com/open?id=0B9IPQTvr2BBkS0VhX0xmS1c5aFk)
 3. `ENV=NAME_OF_ENV ./run-atari.py --load "$ENV".tfmodel --env "$ENV"`
 
-Models are available for the following gym atari environments:
+Models are available for the following gym atari environments (click links for videos):
 
-+ [Breakout-v0](https://gym.openai.com/envs/Breakout-v0)
-+ [AirRaid-v0](https://gym.openai.com/envs/AirRaid-v0)
-+ [Asterix-v0](https://gym.openai.com/envs/Asterix-v0)
-+ [Amidar-v0](https://gym.openai.com/envs/Asterix-v0)
-+ [Seaquest-v0](https://gym.openai.com/envs/Seaquest-v0)
++ [Breakout-v0](https://gym.openai.com/evaluations/eval_L55gczPrQJamMGihq9tzA)
++ [AirRaid-v0](https://gym.openai.com/evaluations/eval_zIeNk5MxSGOmvGEUxrZDUw)
++ [Asterix-v0](https://gym.openai.com/evaluations/eval_mees2c58QfKm5GspCjRfCA)
++ [Amidar-v0](https://gym.openai.com/evaluations/eval_HwEazbHtTYGpCialv9uPhA)
++ [Seaquest-v0](https://gym.openai.com/evaluations/eval_N2624y3NSJWrOgoMSpOi4w)
 
 Note that atari game settings in gym is more difficult than the settings DeepMind papers, therefore the scores are not comparable.

tensorpack/callbacks/param.py

@@ -221,6 +221,7 @@ class StatMonitorParamSetter(HyperParamSetter):
                 cnt += 1
         if cnt >= self.min_non_decreasing \
                 and self.less_than(hist[0], hist[-1]):
+            self.last_changed_epoch = self.epoch_num
             return self.value_func(self.get_current_value())
         return None