add recurrent actor and critic

test/continuous/net.py

@@ -71,16 +71,74 @@ class Critic(nn.Module):
         self.model += [nn.Linear(128, 1)]
         self.model = nn.Sequential(*self.model)
 
+    def forward(self, s, a=None, **kwargs):
+        if not isinstance(s, torch.Tensor):
+            s = torch.tensor(s, device=self.device, dtype=torch.float)
+        batch = s.shape[0]
+        s = s.view(batch, -1)
+        if a is not None:
+            if not isinstance(a, torch.Tensor):
+                a = torch.tensor(a, device=self.device, dtype=torch.float)
+            a = a.view(batch, -1)
+            s = torch.cat([s, a], dim=1)
+        logits = self.model(s)
+        return logits
+
+
+class RecurrentActorProb(nn.Module):
+    def __init__(self, layer_num, state_shape, action_shape,
+                 max_action, device='cpu'):
+        super().__init__()
+        self.device = device
+        self.nn = nn.LSTM(input_size=np.prod(state_shape), hidden_size=128,
+                          num_layers=layer_num, batch_first=True)
+        self.mu = nn.Linear(128, np.prod(action_shape))
+        self.sigma = nn.Parameter(torch.zeros(np.prod(action_shape), 1))
+
+    def forward(self, s, **kwargs):
+        if not isinstance(s, torch.Tensor):
+            s = torch.tensor(s, device=self.device, dtype=torch.float)
+        # s [bsz, len, dim] (training) or [bsz, dim] (evaluation)
+        # In short, the tensor's shape in training phase is longer than which
+        # in evaluation phase.
+        if len(s.shape) == 2:
+            bsz, dim = s.shape
+            length = 1
+        else:
+            bsz, length, dim = s.shape
+        s = s.view(bsz, length, -1)
+        logits, _ = self.nn(s)
+        logits = logits[:, -1]
+        mu = self.mu(logits)
+        shape = [1] * len(mu.shape)
+        shape[1] = -1
+        sigma = (self.sigma.view(shape) + torch.zeros_like(mu)).exp()
+        return (mu, sigma), None
+
+
+class RecurrentCritic(nn.Module):
+    def __init__(self, layer_num, state_shape, action_shape=0, device='cpu'):
+        super().__init__()
+        self.state_shape = state_shape
+        self.action_shape = action_shape
+        self.device = device
+        self.nn = nn.LSTM(input_size=np.prod(state_shape), hidden_size=128,
+                          num_layers=layer_num, batch_first=True)
+        self.fc2 = nn.Linear(128 + np.prod(action_shape), 1)
+
     def forward(self, s, a=None):
         if not isinstance(s, torch.Tensor):
             s = torch.tensor(s, device=self.device, dtype=torch.float)
-        if a is not None and not isinstance(a, torch.Tensor):
-            a = torch.tensor(a, device=self.device, dtype=torch.float)
-        batch = s.shape[0]
-        s = s.view(batch, -1)
-        if a is None:
-            logits = self.model(s)
-        else:
-            a = a.view(batch, -1)
-            logits = self.model(torch.cat([s, a], dim=1))
-        return logits
+        # s [bsz, len, dim] (training) or [bsz, dim] (evaluation)
+        # In short, the tensor's shape in training phase is longer than which
+        # in evaluation phase.
+        assert len(s.shape) == 3
+        self.nn.flatten_parameters()
+        s, (h, c) = self.nn(s)
+        s = s[:, -1]
+        if a is not None:
+            if not isinstance(a, torch.Tensor):
+                a = torch.tensor(a, device=self.device, dtype=torch.float)
+            s = torch.cat([s, a], dim=1)
+        s = self.fc2(s)
+        return s

test/discrete/net.py

@@ -47,8 +47,8 @@ class Critic(nn.Module):
         self.preprocess = preprocess_net
         self.last = nn.Linear(128, 1)
 
-    def forward(self, s):
-        logits, h = self.preprocess(s, None)
+    def forward(self, s, **kwargs):
+        logits, h = self.preprocess(s, state=kwargs.get('state', None))
         logits = self.last(logits)
         return logits
 
@@ -85,7 +85,7 @@ class Recurrent(nn.Module):
             # but pytorch rnn needs [len, bsz, ...]
             s, (h, c) = self.nn(s, (state['h'].transpose(0, 1).contiguous(),
                                     state['c'].transpose(0, 1).contiguous()))
-        s = self.fc2(s)[:, -1]
+        s = self.fc2(s[:, -1])
         # please ensure the first dim is batch size: [bsz, len, ...]
         return s, {'h': h.transpose(0, 1).detach(),
                    'c': c.transpose(0, 1).detach()}

tianshou/data/buffer.py

@@ -5,7 +5,7 @@ from tianshou.data.batch import Batch
 
 class ReplayBuffer(object):
     """:class:`~tianshou.data.ReplayBuffer` stores data generated from
-    interaction between the policy and environment. It stores basically 6 types
+    interaction between the policy and environment. It stores basically 7 types
     of data, as mentioned in :class:`~tianshou.data.Batch`, based on
     ``numpy.ndarray``. Here is the usage:
     ::
@@ -282,6 +282,7 @@ class ReplayBuffer(object):
         return Batch(
             obs=self.get(index, 'obs'),
             act=self.act[index],
+            # act_=self.get(index, 'act'),  # stacked action, for RNN
             rew=self.rew[index],
             done=self.done[index],
             obs_next=self.get(index, 'obs_next'),
@@ -405,6 +406,7 @@ class PrioritizedReplayBuffer(ReplayBuffer):
         return Batch(
             obs=self.get(index, 'obs'),
             act=self.act[index],
+            # act_=self.get(index, 'act'),  # stacked action, for RNN
             rew=self.rew[index],
             done=self.done[index],
             obs_next=self.get(index, 'obs_next'),