Add warnings for duplicate usage of action-bounded actor and action scaling method (#850)

- Fix the current bug discussed in #844 in `test_ppo.py`.
- Add warning for `ActorProb ` if both `max_action ` and
`unbounded=True` are used for model initializations.
- Add warning for PGpolicy and DDPGpolicy if they find duplicate usage
of action-bounded actor and action scaling method.

examples/box2d/bipedal_hardcore_sac.py

@@ -94,13 +94,8 @@ def test_sac_bipedal(args=get_args()):
 
     # model
     net_a = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device)
-    actor = ActorProb(
-        net_a,
-        args.action_shape,
-        max_action=args.max_action,
-        device=args.device,
-        unbounded=True
-    ).to(args.device)
+    actor = ActorProb(net_a, args.action_shape, device=args.device,
+                      unbounded=True).to(args.device)
     actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
 
     net_c1 = Net(

examples/box2d/mcc_sac.py

@@ -67,13 +67,8 @@ def test_sac(args=get_args()):
     test_envs.seed(args.seed)
     # model
     net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device)
-    actor = ActorProb(
-        net,
-        args.action_shape,
-        max_action=args.max_action,
-        device=args.device,
-        unbounded=True
-    ).to(args.device)
+    actor = ActorProb(net, args.action_shape, device=args.device,
+                      unbounded=True).to(args.device)
     actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
     net_c1 = Net(
         args.state_shape,

examples/inverse/irl_gail.py

@@ -119,13 +119,8 @@ def test_gail(args=get_args()):
         activation=nn.Tanh,
         device=args.device
     )
-    actor = ActorProb(
-        net_a,
-        args.action_shape,
-        max_action=args.max_action,
-        unbounded=True,
-        device=args.device
-    ).to(args.device)
+    actor = ActorProb(net_a, args.action_shape, unbounded=True,
+                      device=args.device).to(args.device)
     net_c = Net(
         args.state_shape,
         hidden_sizes=args.hidden_sizes,

examples/mujoco/mujoco_a2c.py

@@ -91,7 +91,6 @@ def test_a2c(args=get_args()):
     actor = ActorProb(
         net_a,
         args.action_shape,
-        max_action=args.max_action,
         unbounded=True,
         device=args.device,
     ).to(args.device)

examples/mujoco/mujoco_npg.py

@@ -93,7 +93,6 @@ def test_npg(args=get_args()):
     actor = ActorProb(
         net_a,
         args.action_shape,
-        max_action=args.max_action,
         unbounded=True,
         device=args.device,
     ).to(args.device)

examples/mujoco/mujoco_ppo.py

@@ -96,7 +96,6 @@ def test_ppo(args=get_args()):
     actor = ActorProb(
         net_a,
         args.action_shape,
-        max_action=args.max_action,
         unbounded=True,
         device=args.device,
     ).to(args.device)

examples/mujoco/mujoco_redq.py

@@ -86,7 +86,6 @@ def test_redq(args=get_args()):
     actor = ActorProb(
         net_a,
         args.action_shape,
-        max_action=args.max_action,
         device=args.device,
         unbounded=True,
         conditioned_sigma=True,

examples/mujoco/mujoco_reinforce.py

@@ -88,7 +88,6 @@ def test_reinforce(args=get_args()):
     actor = ActorProb(
         net_a,
         args.action_shape,
-        max_action=args.max_action,
         unbounded=True,
         device=args.device,
     ).to(args.device)

examples/mujoco/mujoco_sac.py

@@ -81,7 +81,6 @@ def test_sac(args=get_args()):
     actor = ActorProb(
         net_a,
         args.action_shape,
-        max_action=args.max_action,
         device=args.device,
         unbounded=True,
         conditioned_sigma=True,

examples/mujoco/mujoco_trpo.py

@@ -96,7 +96,6 @@ def test_trpo(args=get_args()):
     actor = ActorProb(
         net_a,
         args.action_shape,
-        max_action=args.max_action,
         unbounded=True,
         device=args.device,
     ).to(args.device)

examples/offline/d4rl_cql.py

@@ -108,7 +108,6 @@ def test_cql():
     actor = ActorProb(
         net_a,
         action_shape=args.action_shape,
-        max_action=args.max_action,
         device=args.device,
         unbounded=True,
         conditioned_sigma=True

test/continuous/test_npg.py

@@ -82,13 +82,8 @@ def test_npg(args=get_args()):
         activation=nn.Tanh,
         device=args.device
     )
-    actor = ActorProb(
-        net,
-        args.action_shape,
-        max_action=args.max_action,
-        unbounded=True,
-        device=args.device
-    ).to(args.device)
+    actor = ActorProb(net, args.action_shape, unbounded=True,
+                      device=args.device).to(args.device)
     critic = Critic(
         Net(
             args.state_shape,

test/continuous/test_ppo.py

@@ -81,9 +81,8 @@ def test_ppo(args=get_args()):
     test_envs.seed(args.seed)
     # model
     net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device)
-    actor = ActorProb(
-        net, args.action_shape, max_action=args.max_action, device=args.device
-    ).to(args.device)
+    actor = ActorProb(net, args.action_shape, unbounded=True,
+                      device=args.device).to(args.device)
     critic = Critic(
         Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device),
         device=args.device

test/continuous/test_redq.py

@@ -82,7 +82,6 @@ def test_redq(args=get_args()):
     actor = ActorProb(
         net,
         args.action_shape,
-        max_action=args.max_action,
         device=args.device,
         unbounded=True,
         conditioned_sigma=True

test/continuous/test_sac_with_il.py

@@ -79,13 +79,8 @@ def test_sac_with_il(args=get_args()):
     torch.manual_seed(args.seed)
     # model
     net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device)
-    actor = ActorProb(
-        net,
-        args.action_shape,
-        max_action=args.max_action,
-        device=args.device,
-        unbounded=True
-    ).to(args.device)
+    actor = ActorProb(net, args.action_shape, device=args.device,
+                      unbounded=True).to(args.device)
     actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
     net_c1 = Net(
         args.state_shape,

test/continuous/test_trpo.py

@@ -85,13 +85,8 @@ def test_trpo(args=get_args()):
         activation=nn.Tanh,
         device=args.device
     )
-    actor = ActorProb(
-        net,
-        args.action_shape,
-        max_action=args.max_action,
-        unbounded=True,
-        device=args.device
-    ).to(args.device)
+    actor = ActorProb(net, args.action_shape, unbounded=True,
+                      device=args.device).to(args.device)
     critic = Critic(
         Net(
             args.state_shape,

test/offline/gather_pendulum_data.py

@@ -93,7 +93,6 @@ def gather_data():
     actor = ActorProb(
         net,
         args.action_shape,
-        max_action=args.max_action,
         device=args.device,
         unbounded=True,
     ).to(args.device)

test/offline/test_cql.py

@@ -108,7 +108,6 @@ def test_cql(args=get_args()):
     actor = ActorProb(
         net_a,
         action_shape=args.action_shape,
-        max_action=args.max_action,
         device=args.device,
         unbounded=True,
         conditioned_sigma=True,

tianshou/policy/modelfree/ddpg.py

@@ -64,6 +64,19 @@ class DDPGPolicy(BasePolicy):
         assert action_bound_method != "tanh", "tanh mapping is not supported" \
             "in policies where action is used as input of critic , because" \
             "raw action in range (-inf, inf) will cause instability in training"
+        try:
+            if actor is not None and action_scaling and \
+                    not np.isclose(actor.max_action, 1.):  # type: ignore
+                import warnings
+                warnings.warn(
+                    "action_scaling and action_bound_method are only intended to deal"
+                    "with unbounded model action space, but find actor model bound"
+                    f"action space with max_action={actor.max_action}."
+                    "Consider using unbounded=True option of the actor model,"
+                    "or set action_scaling to False and action_bound_method to \"\"."
+                )
+        except Exception:
+            pass
         if actor is not None and actor_optim is not None:
             self.actor: torch.nn.Module = actor
             self.actor_old = deepcopy(actor)

tianshou/policy/modelfree/pg.py

@@ -53,6 +53,18 @@ class PGPolicy(BasePolicy):
             **kwargs
         )
         self.actor = model
+        try:
+            if action_scaling and not np.isclose(model.max_action, 1.):  # type: ignore
+                import warnings
+                warnings.warn(
+                    "action_scaling and action_bound_method are only intended"
+                    "to deal with unbounded model action space, but find actor model"
+                    f"bound action space with max_action={model.max_action}."
+                    "Consider using unbounded=True option of the actor model,"
+                    "or set action_scaling to False and action_bound_method to \"\"."
+                )
+        except Exception:
+            pass
         self.optim = optim
         self.dist_fn = dist_fn
         assert 0.0 <= discount_factor <= 1.0, "discount factor should be in [0, 1]"

tianshou/utils/net/continuous.py

@@ -1,3 +1,4 @@
+import warnings
 from typing import Any, Dict, Optional, Sequence, Tuple, Type, Union
 
 import numpy as np
@@ -54,7 +55,7 @@ class Actor(nn.Module):
             hidden_sizes,
             device=self.device
         )
-        self._max = max_action
+        self.max_action = max_action
 
     def forward(
         self,
@@ -64,7 +65,7 @@ class Actor(nn.Module):
     ) -> Tuple[torch.Tensor, Any]:
         """Mapping: obs -> logits -> action."""
         logits, hidden = self.preprocess(obs, state)
-        logits = self._max * torch.tanh(self.last(logits))
+        logits = self.max_action * torch.tanh(self.last(logits))
         return logits, hidden
 
 
@@ -178,6 +179,11 @@ class ActorProb(nn.Module):
         preprocess_net_output_dim: Optional[int] = None,
     ) -> None:
         super().__init__()
+        if unbounded and not np.isclose(max_action, 1.0):
+            warnings.warn(
+                "Note that max_action input will be discarded when unbounded is True."
+            )
+            max_action = 1.0
         self.preprocess = preprocess_net
         self.device = device
         self.output_dim = int(np.prod(action_shape))
@@ -198,7 +204,7 @@ class ActorProb(nn.Module):
             )
         else:
             self.sigma_param = nn.Parameter(torch.zeros(self.output_dim, 1))
-        self._max = max_action
+        self.max_action = max_action
         self._unbounded = unbounded
 
     def forward(
@@ -211,7 +217,7 @@ class ActorProb(nn.Module):
         logits, hidden = self.preprocess(obs, state)
         mu = self.mu(logits)
         if not self._unbounded:
-            mu = self._max * torch.tanh(mu)
+            mu = self.max_action * torch.tanh(mu)
         if self._c_sigma:
             sigma = torch.clamp(self.sigma(logits), min=SIGMA_MIN, max=SIGMA_MAX).exp()
         else:
@@ -240,6 +246,11 @@ class RecurrentActorProb(nn.Module):
         conditioned_sigma: bool = False,
     ) -> None:
         super().__init__()
+        if unbounded and not np.isclose(max_action, 1.0):
+            warnings.warn(
+                "Note that max_action input will be discarded when unbounded is True."
+            )
+            max_action = 1.0
         self.device = device
         self.nn = nn.LSTM(
             input_size=int(np.prod(state_shape)),
@@ -254,7 +265,7 @@ class RecurrentActorProb(nn.Module):
             self.sigma = nn.Linear(hidden_layer_size, output_dim)
         else:
             self.sigma_param = nn.Parameter(torch.zeros(output_dim, 1))
-        self._max = max_action
+        self.max_action = max_action
         self._unbounded = unbounded
 
     def forward(
@@ -289,7 +300,7 @@ class RecurrentActorProb(nn.Module):
         logits = obs[:, -1]
         mu = self.mu(logits)
         if not self._unbounded:
-            mu = self._max * torch.tanh(mu)
+            mu = self.max_action * torch.tanh(mu)
         if self._c_sigma:
             sigma = torch.clamp(self.sigma(logits), min=SIGMA_MIN, max=SIGMA_MAX).exp()
         else: