init

pytorch · vmoens · May 23, 2023 · May 23, 2023 · May 23, 2023 · May 23, 2023
commit bc295a0cc47173e254e9b2f37e422080cb975326
diff --git a/examples/iql/iql_online.py b/examples/iql/iql_online.py
@@ -200,7 +200,7 @@ def env_factory(num_workers):
     loss_module.make_value_estimator(gamma=cfg.gamma)
 
     # Define Target Network Updater
-    target_net_updater = SoftUpdate(loss_module, cfg.target_update_polyak)
+    target_net_updater = SoftUpdate(loss_module, eps=cfg.target_update_polyak)
 
     # Make Off-Policy Collector
     collector = SyncDataCollector(

diff --git a/test/test_cost.py b/test/test_cost.py
@@ -3709,15 +3709,19 @@ def __init__(self):
         RuntimeError, match="Your module seems to have a target tensor list "
     ):
         if mode == "hard":
-            upd = HardUpdate(module, value_network_update_interval)
+            upd = HardUpdate(
+                module, value_network_update_interval=value_network_update_interval
+            )
         elif mode == "soft":
-            upd = SoftUpdate(module, 1 - 1 / value_network_update_interval)
+            upd = SoftUpdate(module, eps=1 - 1 / value_network_update_interval)
 
     class custom_module(LossModule):
-        def __init__(self):
+        def __init__(self, delay_module=True):
             super().__init__()
             module1 = torch.nn.BatchNorm2d(10).eval()
-            self.convert_to_functional(module1, "module1", create_target_params=True)
+            self.convert_to_functional(
+                module1, "module1", create_target_params=delay_module
+            )
             module2 = torch.nn.BatchNorm2d(10).eval()
             self.module2 = module2
             iterator_params = self.target_module1_params.values(
@@ -3729,15 +3733,38 @@ def __init__(self):
                 else:
                     target.data += 10
 
+    module = custom_module(delay_module=False)
+    with pytest.raises(RuntimeError, match="The target and source data are identical"):
+        if mode == "hard":
+            upd = HardUpdate(
+                module, value_network_update_interval=value_network_update_interval
+            )
+        elif mode == "soft":
+            upd = SoftUpdate(
+                module,
+                eps=1 - 1 / value_network_update_interval,
+            )
+        else:
+            raise NotImplementedError
+
     module = custom_module().to(device).to(dtype)
+
+    if mode == "soft":
+        with pytest.raises(ValueError, match="One and only one argument"):
+            upd = SoftUpdate(
+                module,
+                eps=1 - 1 / value_network_update_interval,
+                tau=0.1,
+            )
+
     _ = module.module1_params
     _ = module.target_module1_params
     if mode == "hard":
         upd = HardUpdate(
             module, value_network_update_interval=value_network_update_interval
         )
     elif mode == "soft":
-        upd = SoftUpdate(module, 1 - 1 / value_network_update_interval)
+        upd = SoftUpdate(module, eps=1 - 1 / value_network_update_interval)
     for _, _v in upd._targets.items(True, True):
         if _v.dtype is not torch.int64:
             _v.copy_(torch.randn_like(_v))
@@ -5346,14 +5373,20 @@ def fun(a, b, time_dim=-2):
         assert (z2 == 2).all()
 
 
-@pytest.mark.parametrize("updater", [HardUpdate, SoftUpdate])
-def test_updater_warning(updater):
+@pytest.mark.parametrize(
+    "updater,kwarg",
+    [
+        (HardUpdate, {"value_network_update_interval": 1000}),
+        (SoftUpdate, {"eps": 0.99}),
+    ],
+)
+def test_updater_warning(updater, kwarg):
     with warnings.catch_warnings():
         dqn = DQNLoss(torch.nn.Linear(3, 4), delay_value=True, action_space="one_hot")
     with pytest.warns(UserWarning):
         dqn.target_value_network_params
     with warnings.catch_warnings():
-        updater(dqn)
+        updater(dqn, **kwarg)
     with warnings.catch_warnings():
         dqn.target_value_network_params
 

diff --git a/torchrl/objectives/utils.py b/torchrl/objectives/utils.py
@@ -270,14 +270,17 @@ class SoftUpdate(TargetNetUpdater):
 
     This was proposed in "CONTINUOUS CONTROL WITH DEEP REINFORCEMENT LEARNING", https://arxiv.org/pdf/1509.02971.pdf
 
+    One and only one decay factor (tau or eps) must be specified.
+
     Args:
         loss_module (DQNLoss or DDPGLoss): loss module where the target network should be updated.
         eps (scalar): epsilon in the update equation:
             .. math::
 
                 \theta_t = \theta_{t-1} * \epsilon + \theta_t * (1-\epsilon)
 
-            Defaults to 0.999
+            Exclusive with ``tau``.
+        tau (scalar): Polyak tau. It is equal to ``1-eps``, and exclusive with it.
     """
 
     def __init__(
@@ -289,8 +292,27 @@ def __init__(
             "REDQLoss",  # noqa: F821
             "TD3Loss",  # noqa: F821
         ],
-        eps: float = 0.999,
+        *,
+        eps: float = None,
+        tau: Optional[float] = None,
     ):
+        if eps is None and tau is None:
+            warnings.warn(
+                "Neither eps nor tau was provided. Taking the default value "
+                "eps=0.999. This behaviour will soon be deprecated.",
+                category=DeprecationWarning,
+            )
+            eps = 0.999
+        if (eps is None) ^ (tau is None):
+            if eps is None:
+                eps = 1 - tau
+        else:
+            raise ValueError("One and only one argument (tau or eps) can be specified.")
+        if eps < 0.5:
+            warnings.warn(
+                "Found an eps value < 0.5, which is unexpected. "
+                "You may want to use the `tau` keyword argument instead."
+            )
         if not (eps <= 1.0 and eps >= 0.0):
             raise ValueError(
                 f"Got eps = {eps} when it was supposed to be between 0 and 1."
@@ -310,13 +332,16 @@ class HardUpdate(TargetNetUpdater):
 
     Args:
         loss_module (DQNLoss or DDPGLoss): loss module where the target network should be updated.
+
+    Keyword Args:
         value_network_update_interval (scalar): how often the target network should be updated.
             default: 1000
     """
 
     def __init__(
         self,
         loss_module: Union["DQNLoss", "DDPGLoss", "SACLoss", "TD3Loss"],  # noqa: F821
+        *,
         value_network_update_interval: float = 1000,
     ):
         super(HardUpdate, self).__init__(loss_module)

diff --git a/torchrl/trainers/helpers/losses.py b/torchrl/trainers/helpers/losses.py
@@ -34,11 +34,12 @@ def make_target_updater(
     if cfg.loss == "double":
         if not cfg.hard_update:
             target_net_updater = SoftUpdate(
-                loss_module, 1 - 1 / cfg.value_network_update_interval
+                loss_module, eps=1 - 1 / cfg.value_network_update_interval
             )
         else:
             target_net_updater = HardUpdate(
-                loss_module, cfg.value_network_update_interval
+                loss_module,
+                value_network_update_interval=cfg.value_network_update_interval,
             )
     else:
         if cfg.hard_update:

diff --git a/tutorials/sphinx-tutorials/coding_dqn.py b/tutorials/sphinx-tutorials/coding_dqn.py
@@ -440,7 +440,7 @@ def get_collector(
 def get_loss_module(actor, gamma):
     loss_module = DQNLoss(actor, delay_value=True)
     loss_module.make_value_estimator(gamma=gamma)
-    target_updater = SoftUpdate(loss_module)
+    target_updater = SoftUpdate(loss_module, eps=0.995)
     return loss_module, target_updater