[Feature] ConditionalPolicySwitch transform

ghstack-source-id: defb61a
Pull Request resolved: #2711

Author: vmoens

docs/source/reference/envs.rst

@@ -816,6 +816,7 @@ to be able to create this other composition:
     CenterCrop
     ClipTransform
     Compose
+    ConditionalPolicySwitch
     Crop
     DTypeCastTransform
     DeviceCastTransform

examples/agents/ppo-chess.py

@@ -5,20 +5,24 @@
 import tensordict.nn
 import torch
 import tqdm
-from tensordict.nn import TensorDictSequential as TDSeq, TensorDictModule as TDMod, \
-    ProbabilisticTensorDictModule as TDProb, ProbabilisticTensorDictSequential as TDProbSeq
+from tensordict.nn import (
+    ProbabilisticTensorDictModule as TDProb,
+    ProbabilisticTensorDictSequential as TDProbSeq,
+    TensorDictModule as TDMod,
+    TensorDictSequential as TDSeq,
+)
 from torch import nn
 from torch.nn.utils import clip_grad_norm_
 from torch.optim import Adam
 
 from torchrl.collectors import SyncDataCollector
+from torchrl.data import LazyTensorStorage, ReplayBuffer, SamplerWithoutReplacement
 
 from torchrl.envs import ChessEnv, Tokenizer
 from torchrl.modules import MLP
 from torchrl.modules.distributions import MaskedCategorical
 from torchrl.objectives import ClipPPOLoss
 from torchrl.objectives.value import GAE
-from torchrl.data import ReplayBuffer, LazyTensorStorage, SamplerWithoutReplacement
 
 tensordict.nn.set_composite_lp_aggregate(False)
 
@@ -39,7 +43,9 @@
 embedding_moves = nn.Embedding(num_embeddings=n + 1, embedding_dim=64)
 
 # Embedding for the fen
-embedding_fen = nn.Embedding(num_embeddings=transform.tokenizer.vocab_size, embedding_dim=64)
+embedding_fen = nn.Embedding(
+    num_embeddings=transform.tokenizer.vocab_size, embedding_dim=64
+)
 
 backbone = MLP(out_features=512, num_cells=[512] * 8, activation_class=nn.ReLU)
 
@@ -49,20 +55,30 @@
 critic_head = nn.Linear(512, 1)
 critic_head.bias.data.fill_(0)
 
-prob = TDProb(in_keys=["logits", "mask"], out_keys=["action"], distribution_class=MaskedCategorical, return_log_prob=True)
+prob = TDProb(
+    in_keys=["logits", "mask"],
+    out_keys=["action"],
+    distribution_class=MaskedCategorical,
+    return_log_prob=True,
+)
+
 
 def make_mask(idx):
     mask = idx.new_zeros((*idx.shape[:-1], n + 1), dtype=torch.bool)
     return mask.scatter_(-1, idx, torch.ones_like(idx, dtype=torch.bool))[..., :-1]
 
+
 actor = TDProbSeq(
-    TDMod(
-        make_mask,
-        in_keys=["legal_moves"], out_keys=["mask"]),
+    TDMod(make_mask, in_keys=["legal_moves"], out_keys=["mask"]),
     TDMod(embedding_moves, in_keys=["legal_moves"], out_keys=["embedded_legal_moves"]),
     TDMod(embedding_fen, in_keys=["fen_tokenized"], out_keys=["embedded_fen"]),
-    TDMod(lambda *args: torch.cat([arg.view(*arg.shape[:-2], -1) for arg in args], dim=-1), in_keys=["embedded_legal_moves", "embedded_fen"],
-          out_keys=["features"]),
+    TDMod(
+        lambda *args: torch.cat(
+            [arg.view(*arg.shape[:-2], -1) for arg in args], dim=-1
+        ),
+        in_keys=["embedded_legal_moves", "embedded_fen"],
+        out_keys=["features"],
+    ),
     TDMod(backbone, in_keys=["features"], out_keys=["hidden"]),
     TDMod(actor_head, in_keys=["hidden"], out_keys=["logits"]),
     prob,
@@ -78,7 +94,9 @@ def make_mask(idx):
 
 optim = Adam(loss.parameters())
 
-gae = GAE(value_network=TDSeq(*actor[:-2], critic), gamma=0.99, lmbda=0.95, shifted=True)
+gae = GAE(
+    value_network=TDSeq(*actor[:-2], critic), gamma=0.99, lmbda=0.95, shifted=True
+)
 
 # Create a data collector
 collector = SyncDataCollector(
@@ -88,12 +106,20 @@ def make_mask(idx):
     total_frames=1_000_000,
 )
 
-replay_buffer0 = ReplayBuffer(storage=LazyTensorStorage(max_size=collector.frames_per_batch//2), batch_size=batch_size, sampler=SamplerWithoutReplacement())
-replay_buffer1 = ReplayBuffer(storage=LazyTensorStorage(max_size=collector.frames_per_batch//2), batch_size=batch_size, sampler=SamplerWithoutReplacement())
+replay_buffer0 = ReplayBuffer(
+    storage=LazyTensorStorage(max_size=collector.frames_per_batch // 2),
+    batch_size=batch_size,
+    sampler=SamplerWithoutReplacement(),
+)
+replay_buffer1 = ReplayBuffer(
+    storage=LazyTensorStorage(max_size=collector.frames_per_batch // 2),
+    batch_size=batch_size,
+    sampler=SamplerWithoutReplacement(),
+)
 
 for data in tqdm.tqdm(collector):
     data = data.filter_non_tensor_data()
-    print('data', data[0::2])
+    print("data", data[0::2])
     for i in range(num_epochs):
         replay_buffer0.empty()
         replay_buffer1.empty()
@@ -103,14 +129,24 @@ def make_mask(idx):
             # player 1
             data1 = gae(data[1::2])
             if i == 0:
-                print('win rate for 0', data0["next", "reward"].sum() / data["next", "done"].sum().clamp_min(1e-6))
-                print('win rate for 1', data1["next", "reward"].sum() / data["next", "done"].sum().clamp_min(1e-6))
+                print(
+                    "win rate for 0",
+                    data0["next", "reward"].sum()
+                    / data["next", "done"].sum().clamp_min(1e-6),
+                )
+                print(
+                    "win rate for 1",
+                    data1["next", "reward"].sum()
+                    / data["next", "done"].sum().clamp_min(1e-6),
+                )
 
             replay_buffer0.extend(data0)
             replay_buffer1.extend(data1)
 
-        n_iter = collector.frames_per_batch//(2 * batch_size)
-        for (d0, d1) in tqdm.tqdm(zip(replay_buffer0, replay_buffer1, strict=True), total=n_iter):
+        n_iter = collector.frames_per_batch // (2 * batch_size)
+        for (d0, d1) in tqdm.tqdm(
+            zip(replay_buffer0, replay_buffer1, strict=True), total=n_iter
+        ):
             loss_vals = (loss(d0) + loss(d1)) / 2
             loss_vals.sum(reduce=True).backward()
             gn = clip_grad_norm_(loss.parameters(), 100.0)

test/test_transforms.py

@@ -20,6 +20,7 @@
 
 import tensordict.tensordict
 import torch
+from tensordict.nn import WrapModule
 
 from torchrl.collectors import MultiSyncDataCollector
 
@@ -106,6 +107,7 @@
     CenterCrop,
     ClipTransform,
     Compose,
+    ConditionalPolicySwitch,
     Crop,
     DeviceCastTransform,
     DiscreteActionProjection,
@@ -13192,6 +13194,175 @@ def test_composite_reward_spec(self) -> None:
         assert transform.transform_reward_spec(reward_spec) == expected_reward_spec
 
 
+class TestConditionalPolicySwitch(TransformBase):
+    def test_single_trans_env_check(self):
+        base_env = CountingEnv(max_steps=15)
+        condition = lambda td: ((td.get("step_count") % 2) == 0).all()
+        # Player 0
+        policy_odd = lambda td: td.set("action", env.action_spec.zero())
+        policy_even = lambda td: td.set("action", env.action_spec.one())
+        transforms = Compose(
+            StepCounter(),
+            ConditionalPolicySwitch(condition=condition, policy=policy_even),
+        )
+        env = base_env.append_transform(transforms)
+        r = env.rollout(1000, policy_odd, break_when_all_done=True)
+        assert r.shape[0] == 15
+        assert (r["action"] == 0).all()
+        assert (
+            r["step_count"] == torch.arange(1, r.numel() * 2, 2).unsqueeze(-1)
+        ).all()
+        assert r["next", "done"].any()
+
+        # Player 1
+        condition = lambda td: ((td.get("step_count") % 2) == 1).all()
+        transforms = Compose(
+            StepCounter(),
+            ConditionalPolicySwitch(condition=condition, policy=policy_odd),
+        )
+        env = base_env.append_transform(transforms)
+        r = env.rollout(1000, policy_even, break_when_all_done=True)
+        assert r.shape[0] == 16
+        assert (r["action"] == 1).all()
+        assert (
+            r["step_count"] == torch.arange(0, r.numel() * 2, 2).unsqueeze(-1)
+        ).all()
+        assert r["next", "done"].any()
+
+    def _create_policy_odd(self, base_env):
+        return WrapModule(
+            lambda td, base_env=base_env: td.set(
+                "action", base_env.action_spec_unbatched.zero(td.shape)
+            ),
+            out_keys=["action"],
+        )
+
+    def _create_policy_even(self, base_env):
+        return WrapModule(
+            lambda td, base_env=base_env: td.set(
+                "action", base_env.action_spec_unbatched.one(td.shape)
+            ),
+            out_keys=["action"],
+        )
+
+    def _create_transforms(self, condition, policy_even):
+        return Compose(
+            StepCounter(),
+            ConditionalPolicySwitch(condition=condition, policy=policy_even),
+        )
+
+    def _make_env(self, max_count, env_cls):
+        torch.manual_seed(0)
+        condition = lambda td: ((td.get("step_count") % 2) == 0).squeeze(-1)
+        base_env = env_cls(max_steps=max_count)
+        policy_even = self._create_policy_even(base_env)
+        transforms = self._create_transforms(condition, policy_even)
+        return base_env.append_transform(transforms)
+
+    def _test_env(self, env, policy_odd):
+        env.check_env_specs()
+        env.set_seed(0)
+        r = env.rollout(100, policy_odd, break_when_any_done=False)
+        # Check results are independent: one reset / step in one env should not impact results in another
+        r0, r1, r2 = r.unbind(0)
+        r0_split = r0.split(6)
+        assert all(((r == r0_split[0][: r.numel()]).all() for r in r0_split[1:]))
+        r1_split = r1.split(7)
+        assert all(((r == r1_split[0][: r.numel()]).all() for r in r1_split[1:]))
+        r2_split = r2.split(8)
+        assert all(((r == r2_split[0][: r.numel()]).all() for r in r2_split[1:]))
+
+    def test_trans_serial_env_check(self):
+        torch.manual_seed(0)
+        base_env = SerialEnv(
+            3,
+            [partial(CountingEnv, 6), partial(CountingEnv, 7), partial(CountingEnv, 8)],
+            batch_locked=False,
+        )
+        condition = lambda td: ((td.get("step_count") % 2) == 0).squeeze(-1)
+        policy_odd = self._create_policy_odd(base_env)
+        policy_even = self._create_policy_even(base_env)
+        transforms = self._create_transforms(condition, policy_even)
+        env = base_env.append_transform(transforms)
+        self._test_env(env, policy_odd)
+
+    def test_trans_parallel_env_check(self):
+        torch.manual_seed(0)
+        base_env = ParallelEnv(
+            3,
+            [partial(CountingEnv, 6), partial(CountingEnv, 7), partial(CountingEnv, 8)],
+            batch_locked=False,
+            mp_start_method=mp_ctx,
+        )
+        condition = lambda td: ((td.get("step_count") % 2) == 0).squeeze(-1)
+        policy_odd = self._create_policy_odd(base_env)
+        policy_even = self._create_policy_even(base_env)
+        transforms = self._create_transforms(condition, policy_even)
+        env = base_env.append_transform(transforms)
+        self._test_env(env, policy_odd)
+
+    def test_serial_trans_env_check(self):
+        condition = lambda td: ((td.get("step_count") % 2) == 0).squeeze(-1)
+        policy_odd = self._create_policy_odd(CountingEnv())
+
+        def make_env(max_count):
+            return partial(self._make_env, max_count, CountingEnv)
+
+        env = SerialEnv(3, [make_env(6), make_env(7), make_env(8)])
+        self._test_env(env, policy_odd)
+
+    def test_parallel_trans_env_check(self):
+        condition = lambda td: ((td.get("step_count") % 2) == 0).squeeze(-1)
+        policy_odd = self._create_policy_odd(CountingEnv())
+
+        def make_env(max_count):
+            return partial(self._make_env, max_count, CountingEnv)
+
+        env = ParallelEnv(
+            3, [make_env(6), make_env(7), make_env(8)], mp_start_method=mp_ctx
+        )
+        self._test_env(env, policy_odd)
+
+    def test_transform_no_env(self):
+        """tests the transform on dummy data, without an env."""
+        raise NotImplementedError
+
+    def test_transform_compose(self):
+        """tests the transform on dummy data, without an env but inside a Compose."""
+        raise NotImplementedError
+
+    def test_transform_env(self):
+        """tests the transform on a real env.
+
+        If possible, do not use a mock env, as bugs may go unnoticed if the dynamic is too
+        simplistic. A call to reset() and step() should be tested independently, ie
+        a check that reset produces the desired output and that step() does too.
+
+        """
+        raise NotImplementedError
+
+    def test_transform_model(self):
+        """tests the transform before an nn.Module that reads the output."""
+        raise NotImplementedError
+
+    def test_transform_rb(self):
+        """tests the transform when used with a replay buffer.
+
+        If your transform is not supposed to work with a replay buffer, test that
+        an error will be raised when called or appended to a RB.
+
+        """
+        raise NotImplementedError
+
+    def test_transform_inverse(self):
+        """tests the inverse transform. If not applicable, simply skip this test.
+
+        If your transform is not supposed to work offline, test that
+        an error will be raised when called in a nn.Module.
+        """
+        raise NotImplementedError
+
+
 if __name__ == "__main__":
     args, unknown = argparse.ArgumentParser().parse_known_args()
     pytest.main([__file__, "--capture", "no", "--exitfirst"] + unknown)

torchrl/envs/__init__.py

@@ -55,6 +55,7 @@
     CenterCrop,
     ClipTransform,
     Compose,
+    ConditionalPolicySwitch,
     Crop,
     DeviceCastTransform,
     DiscreteActionProjection,

torchrl/envs/batched_envs.py

@@ -191,6 +191,8 @@ class BatchedEnvBase(EnvBase):
             one of the environment has dynamic specs.
 
               .. note:: Learn more about dynamic specs and environments :ref:`here <dynamic_envs>`.
+        batch_locked (bool, optional): if provided, will override the ``batch_locked`` attribute of the
+            nested environments. `batch_locked=False` may allow for partial steps.
 
     .. note::
         One can pass keyword arguments to each sub-environments using the following
@@ -305,6 +307,7 @@ def __init__(
         non_blocking: bool = False,
         mp_start_method: str = None,
         use_buffers: bool = None,
+        batch_locked: bool | None = None,
     ):
         super().__init__(device=device)
         self.serial_for_single = serial_for_single
@@ -344,6 +347,7 @@ def __init__(
 
         # if share_individual_td is None, we will assess later if the output can be stacked
         self.share_individual_td = share_individual_td
+        self._batch_locked = batch_locked
         self._share_memory = shared_memory
         self._memmap = memmap
         self.allow_step_when_done = allow_step_when_done
@@ -610,8 +614,8 @@ def map_device(key, value, device_map=device_map):
                 self._env_tensordict.named_apply(
                     map_device, nested_keys=True, filter_empty=True
                 )
-
-            self._batch_locked = meta_data.batch_locked
+            if self._batch_locked is None:
+                self._batch_locked = meta_data.batch_locked
         else:
             self._batch_size = torch.Size([self.num_workers, *meta_data[0].batch_size])
             devices = set()
@@ -652,7 +656,8 @@ def map_device(key, value, device_map=device_map):
                 self._env_tensordict = torch.stack(
                     [meta_data.tensordict for meta_data in meta_data], 0
                 )
-            self._batch_locked = meta_data[0].batch_locked
+            if self._batch_locked is None:
+                self._batch_locked = meta_data[0].batch_locked
         self.has_lazy_inputs = contains_lazy_spec(self.input_spec)
 
     def state_dict(self) -> OrderedDict: