update readme

README.md

@@ -12,8 +12,12 @@ See some interesting [examples](https://github.com/ppwwyyxx/tensorpack/tree/mast
 
 ## Features:
 
-Focused on modularity:
+Focused on modularity. Just have to define the three components in training:
 
-+ Models has scoped abstraction of common models.
-+ Dataflow defines data preprocessing in pure Python.
-+ Callbacks systems controls training behavior.
+1. The model, or the graph. Define its input and output. `models/` has some scoped abstraction of common models.
+
+2. The data. All data producer has a unified `DataFlow` interface, and this interface can be chained
+	 to perform complex preprocessing. It uses multiprocess to avoid performance bottleneck.
+
+3. The callbacks. They include everything you want to do besides the training iterations:
+	change hyperparameters, save model, print logs, run validation, and more.

examples/Atari2600/README.md

-Implement DQN in:
+Reproduce DQN in:
 
 **Human-level Control Through Deep Reinforcement Learning**
 
@@ -6,10 +6,26 @@ and Double-DQN in:
 
 **Deep Reinforcement Learning with Double Q-learning**
 
-To run:
+Can reproduce the claimed performance, on several games I've tested with.
+
+![DQN](https://github.com/ppwwyyxx/tensorpack/raw/master/examples/Atari2600/DoubleDQN-breakout.png)
+
+A demo trained with Double-DQN on breakout is available at [youtube](https://youtu.be/o21mddZtE5Y).
+
+## How to use
+
+Download [atari roms](https://github.com/openai/atari-py/tree/master/atari_py/atari_roms) to
+`$TENSORPACK_DATASET/atari_rom` (defaults to tensorpack/dataflow/dataset/atari_rom).
+
+To train:
 ```
-./DQN.py --rom breakout.rom --gpu 0
+./DQN.py --rom breakout.bin --gpu 0
 ```
+Training speed is about 7.3 iteration/s on 1 Tesla M40. It takes days to learn well (see figure above).
 
-Can reproduce the claimed performance, on games I've tested with (curves will be available soon).
-A demo trained with Double-DQN on breakout is available at [youtube](https://youtu.be/o21mddZtE5Y).
+To play:
+```
+./DQN.py --rom breakout.bin --task play --load pretrained.model
+```
+
+A3C code and curve will be available soon. It learns much faster.

examples/ResNet/README.md

@@ -9,5 +9,7 @@ The validation error here is computed on test set.
 
 ![cifar10](https://github.com/ppwwyyxx/tensorpack/raw/master/examples/ResNet/cifar10-resnet.png)
 
-Download model:
-[Cifar10 n=18](https://drive.google.com/open?id=0B308TeQzmFDLeHpSaHAxWGV1WDg)
+<!--
+   -Download model:
+   -[Cifar10 n=18](https://drive.google.com/open?id=0B308TeQzmFDLeHpSaHAxWGV1WDg)
+	 -->

tensorpack/RL/atari.py

@@ -26,6 +26,7 @@ __all__ = ['AtariPlayer']
 def log_once():
     logger.warn("https://github.com/mgbellemare/Arcade-Learning-Environment/pull/171 is not merged!")
 
+ROM_URL = "https://github.com/openai/atari-py/tree/master/atari_py/atari_roms"
 _ALE_LOCK = threading.Lock()
 
 class AtariPlayer(RLEnvironment):
@@ -51,7 +52,8 @@ class AtariPlayer(RLEnvironment):
         super(AtariPlayer, self).__init__()
         if not os.path.isfile(rom_file) and '/' not in rom_file:
             rom_file = os.path.join(get_dataset_dir('atari_rom'), rom_file)
-        assert os.path.isfile(rom_file), "rom {} not found".format(rom_file)
+        assert os.path.isfile(rom_file), \
+                "rom {} not found. Please download at {}".format(rom_file, ROM_URL)
 
         try:
             ALEInterface.setLoggerMode(ALEInterface.Logger.Warning)