[DQN] update docs; make eval configurable

examples/DeepQNetwork/DQN.py

@@ -26,7 +26,6 @@ MEMORY_SIZE = 1e6
 # will consume at least 1e6 * 84 * 84 bytes == 6.6G memory.
 INIT_MEMORY_SIZE = MEMORY_SIZE // 20
 STEPS_PER_EPOCH = 100000 // UPDATE_FREQ  # each epoch is 100k state transitions
-EVAL_EPISODE = 50
 NUM_PARALLEL_PLAYERS = 3
 
 USE_GYM = False
@@ -135,7 +134,7 @@ def get_config(model):
                 [(0, 1), (10, 0.1), (400, 0.01)],   # 1->0.1 in the first million steps
                 interp='linear'),
             PeriodicTrigger(Evaluator(
-                EVAL_EPISODE, ['state'], ['Qvalue'], get_player),
+                args.num_eval, ['state'], ['Qvalue'], get_player),
                 every_k_epochs=5 if 'pong' in args.env.lower() else 10),  # eval more frequently for easy games
         ],
         steps_per_epoch=STEPS_PER_EPOCH,
@@ -153,6 +152,7 @@ if __name__ == '__main__':
                         help='either an atari rom file (that ends with .bin) or a gym atari environment name')
     parser.add_argument('--algo', help='algorithm',
                         choices=['DQN', 'Double', 'Dueling'], default='Double')
+    parser.add_argument('--num-eval', default=50, type=int)
     args = parser.parse_args()
     if args.gpu:
         os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
@@ -177,7 +177,7 @@ if __name__ == '__main__':
         if args.task == 'play':
             play_n_episodes(get_player(viz=0.01), pred, 100, render=True)
         elif args.task == 'eval':
-            eval_model_multithread(pred, EVAL_EPISODE, get_player)
+            eval_model_multithread(pred, args.num_eval, get_player)
     else:
         logger.set_logger_dir(
             os.path.join('train_log', 'DQN-{}'.format(

examples/DeepQNetwork/DQNModel.py

@@ -100,7 +100,7 @@ class Model(ModelDesc):
 
     def optimizer(self):
         lr = tf.get_variable('learning_rate', initializer=1e-3, trainable=False)
-        tf.summary.scalar("learning_rate", lr)
+        tf.summary.scalar("learning_rate-summary", lr)
         opt = tf.train.RMSPropOptimizer(lr, decay=0.95, momentum=0.95, epsilon=1e-2)
         return optimizer.apply_grad_processors(opt, [gradproc.SummaryGradient()])
 

examples/DeepQNetwork/README.md

@@ -54,10 +54,11 @@ Claimed performance in the paper can be reproduced, on several games I've tested
 ![DQN](curve-breakout.png)
 
 | Environment    | Avg Score | Download                                                                               |
-|:--------------|:---------:|:-------------------------------------------------------------------------------------:|
+|:---------------|:---------:|:--------------------------------------------------------------------------------------:|
 | breakout.bin   | 465       | [:arrow_down:](http://models.tensorpack.com/DeepQNetwork/DoubleDQN-breakout.bin.npz)   |
 | seaquest.bin   | 8686      | [:arrow_down:](http://models.tensorpack.com/DeepQNetwork/DoubleDQN-seaquest.bin.npz)   |
 | ms_pacman.bin  | 3323      | [:arrow_down:](http://models.tensorpack.com/DeepQNetwork/DoubleDQN-ms_pacman.bin.npz)  |
+| beam_rider.bin | 15835     | [:arrow_down:](http://models.tensorpack.com/DeepQNetwork/DoubleDQN-beam_rider.bin.npz) |
 
 ## Speed
 On one GTX 1080Ti,