Categorical DQN for Atari

main.py

@@ -355,6 +355,29 @@ def categorical_dqn_cart_pole():
     config.categorical_n_atoms = 50
     run_episodes(CategoricalDQNAgent(config))
 
+def categorical_dqn_pixel_atari(name):
+    config = Config()
+    config.history_length = 4
+    config.task_fn = lambda: PixelAtari(name, no_op=30, frame_skip=4, normalized_state=False,
+                                        history_length=config.history_length)
+    action_dim = config.task_fn().action_dim
+    config.optimizer_fn = lambda params: torch.optim.Adam(params, lr=0.00025, eps=0.01 / 32)
+    config.network_fn = lambda: CategoricalConvNet(config.history_length, action_dim, config.categorical_n_atoms, gpu=0)
+    config.policy_fn = lambda: GreedyPolicy(epsilon=1.0, final_step=1000000, min_epsilon=0.1)
+    config.replay_fn = lambda: Replay(memory_size=1000000, batch_size=32, dtype=np.uint8)
+    config.reward_shift_fn = lambda r: np.sign(r)
+    config.discount = 0.99
+    config.target_network_update_freq = 10000
+    config.exploration_steps= 50000
+    config.logger = Logger('./log', logger)
+    config.test_interval = 10
+    config.test_repetitions = 1
+    config.double_q = False
+    config.categorical_v_max = 10
+    config.categorical_v_min = -10
+    config.categorical_n_atoms = 51
+    run_episodes(CategoricalDQNAgent(config))
+
 if __name__ == '__main__':
     mkdir('data')
     mkdir('data/video')
@@ -364,7 +387,7 @@ if __name__ == '__main__':
     logger.setLevel(logging.INFO)
 
     # dqn_cart_pole()
-    categorical_dqn_cart_pole()
+    # categorical_dqn_cart_pole()
     # async_cart_pole()
     # a3c_cart_pole()
     # a2c_cart_pole()
@@ -374,6 +397,7 @@ if __name__ == '__main__':
     # ddpg_continuous()
 
     # dqn_pixel_atari('PongNoFrameskip-v4')
+    categorical_dqn_pixel_atari('PongNoFrameskip-v4')
     # async_pixel_atari('PongNoFrameskip-v4')
     # a3c_pixel_atari('PongNoFrameskip-v4')
     # a2c_pixel_atari('PongNoFrameskip-v4')

network/conv_network.py

@@ -141,3 +141,24 @@ class NatureActorCriticConvNet(nn.Module, ActorCriticNet):
         x = x.view(x.size(0), -1)
         phi = F.relu(self.fc4(x))
         return phi
+
+class CategoricalConvNet(nn.Module, CategoricalNet):
+    def __init__(self, in_channels, n_actions, n_atoms, gpu=0):
+        super(CategoricalConvNet, self).__init__()
+        self.conv1 = nn.Conv2d(in_channels, 32, kernel_size=8, stride=4)
+        self.conv2 = nn.Conv2d(32, 64, kernel_size=4, stride=2)
+        self.conv3 = nn.Conv2d(64, 64, kernel_size=3, stride=1)
+        self.fc4 = nn.Linear(7 * 7 * 64, 512)
+        self.fc_categorical = nn.Linear(512, n_actions * n_atoms)
+        self.n_actions = n_actions
+        self.n_atoms = n_atoms
+        BasicNet.__init__(self, gpu)
+
+    def forward(self, x):
+        x = self.variable(x)
+        y = F.relu(self.conv1(x))
+        y = F.relu(self.conv2(y))
+        y = F.relu(self.conv3(y))
+        y = y.view(y.size(0), -1)
+        y = F.relu(self.fc4(y))
+        return y