[Feature] multiagent data standardization: PPO advantages (#2677)

Co-authored-by: Vincent Moens <vmoens@meta.com>

Author: matteobettini

test/test_cost.py

@@ -19,7 +19,6 @@
 import torch
 
 from packaging import version, version as pack_version
-
 from tensordict import assert_allclose_td, TensorDict, TensorDictBase
 from tensordict._C import unravel_keys
 from tensordict.nn import (
@@ -38,6 +37,7 @@
 from tensordict.nn.utils import Buffer
 from tensordict.utils import unravel_key
 from torch import autograd, nn
+from torchrl._utils import _standardize
 from torchrl.data import Bounded, Categorical, Composite, MultiOneHot, OneHot, Unbounded
 from torchrl.data.postprocs.postprocs import MultiStep
 from torchrl.envs.model_based.dreamer import DreamerEnv
@@ -15848,6 +15848,18 @@ class _AcceptedKeys:
 
 
 class TestUtils:
+    def test_standardization(self):
+        t = torch.arange(3 * 4 * 5 * 6, dtype=torch.float32).view(3, 4, 5, 6)
+        std_t0 = _standardize(t, exclude_dims=(1, 3))
+        std_t1 = (t - t.mean((0, 2), keepdim=True)) / t.std((0, 2), keepdim=True).clamp(
+            1 - 6
+        )
+        torch.testing.assert_close(std_t0, std_t1)
+        std_t = _standardize(t, (), -1, 2)
+        torch.testing.assert_close(std_t, (t + 1) / 2)
+        std_t = _standardize(t, ())
+        torch.testing.assert_close(std_t, (t - t.mean()) / t.std())
+
     @pytest.mark.parametrize("B", [None, (1, ), (4, ), (2, 2, ), (1, 2, 8, )])  # fmt: skip
     @pytest.mark.parametrize("T", [1, 10])
     @pytest.mark.parametrize("device", get_default_devices())

torchrl/_utils.py

@@ -24,15 +24,15 @@
 from distutils.util import strtobool
 from functools import wraps
 from importlib import import_module
-from typing import Any, Callable, cast, Dict, TypeVar, Union
+from typing import Any, Callable, cast, Dict, Tuple, TypeVar, Union
 
 import numpy as np
 import torch
 from packaging.version import parse
 from tensordict import unravel_key
 
 from tensordict.utils import NestedKey
-from torch import multiprocessing as mp
+from torch import multiprocessing as mp, Tensor
 
 try:
     from torch.compiler import is_compiling
@@ -872,6 +872,70 @@ def set_mode(self, type: Any | None) -> None:
             self._mode = type
 
 
+def _standardize(
+    input: Tensor,
+    exclude_dims: Tuple[int] = (),
+    mean: Tensor | None = None,
+    std: Tensor | None = None,
+    eps: float | None = None,
+):
+    """Standardizes the input tensor with the possibility of excluding specific dims from the statistics.
+
+    Useful when processing multi-agent data to keep the agent dimensions independent.
+
+    Args:
+        input (Tensor): the input tensor to be standardized.
+        exclude_dims (Tuple[int]): dimensions to exclude from the statistics, can be negative. Default: ().
+        mean (Tensor): a mean to be used for standardization. Must be of shape broadcastable to input. Default: None.
+        std (Tensor): a standard deviation to be used for standardization. Must be of shape broadcastable to input. Default: None.
+        eps (float): epsilon to be used for numerical stability. Default: float32 resolution.
+
+    """
+    if eps is None:
+        if input.dtype.is_floating_point:
+            eps = torch.finfo(torch.float).resolution
+        else:
+            eps = 1e-6
+
+    len_exclude_dims = len(exclude_dims)
+    if not len_exclude_dims:
+        if mean is None:
+            mean = input.mean()
+        else:
+            # Assume dtypes are compatible
+            mean = torch.as_tensor(mean, device=input.device)
+        if std is None:
+            std = input.std()
+        else:
+            # Assume dtypes are compatible
+            std = torch.as_tensor(std, device=input.device)
+        return (input - mean) / std.clamp_min(eps)
+
+    input_shape = input.shape
+    exclude_dims = [
+        d if d >= 0 else d + len(input_shape) for d in exclude_dims
+    ]  # Make negative dims positive
+
+    if len(set(exclude_dims)) != len_exclude_dims:
+        raise ValueError("Exclude dims has repeating elements")
+    if any(dim < 0 or dim >= len(input_shape) for dim in exclude_dims):
+        raise ValueError(
+            f"exclude_dims={exclude_dims} provided outside bounds for input of shape={input_shape}"
+        )
+    if len_exclude_dims == len(input_shape):
+        warnings.warn(
+            "_standardize called but all dims were excluded from the statistics, returning unprocessed input"
+        )
+        return input
+
+    included_dims = tuple(d for d in range(len(input_shape)) if d not in exclude_dims)
+    if mean is None:
+        mean = torch.mean(input, keepdim=True, dim=included_dims)
+    if std is None:
+        std = torch.std(input, keepdim=True, dim=included_dims)
+    return (input - mean) / std.clamp_min(eps)
+
+
 @wraps(torch.compile)
 def compile_with_warmup(*args, warmup: int = 1, **kwargs):
     """Compile a model with warm-up.

torchrl/objectives/ppo.py

@@ -27,6 +27,7 @@
 from tensordict.utils import NestedKey
 from torch import distributions as d
 
+from torchrl._utils import _standardize
 from torchrl.objectives.common import LossModule
 
 from torchrl.objectives.utils import (
@@ -46,6 +47,7 @@
     TDLambdaEstimator,
     VTrace,
 )
+from yaml import warnings
 
 
 class PPOLoss(LossModule):
@@ -87,6 +89,9 @@ class PPOLoss(LossModule):
             Can be one of "l1", "l2" or "smooth_l1". Defaults to ``"smooth_l1"``.
         normalize_advantage (bool, optional): if ``True``, the advantage will be normalized
             before being used. Defaults to ``False``.
+        normalize_advantage_exclude_dims (Tuple[int], optional): dimensions to exclude from the advantage
+            standardization. Negative dimensions are valid. This is useful in multiagent (or multiobjective) settings
+            where the agent (or objective) dimension may be excluded from the reductions. Default: ().
         separate_losses (bool, optional): if ``True``, shared parameters between
             policy and critic will only be trained on the policy loss.
             Defaults to ``False``, i.e., gradients are propagated to shared
@@ -311,6 +316,7 @@ def __init__(
         critic_coef: float = 1.0,
         loss_critic_type: str = "smooth_l1",
         normalize_advantage: bool = False,
+        normalize_advantage_exclude_dims: Tuple[int] = (),
         gamma: float = None,
         separate_losses: bool = False,
         advantage_key: str = None,
@@ -381,6 +387,8 @@ def __init__(
             self.critic_coef = None
         self.loss_critic_type = loss_critic_type
         self.normalize_advantage = normalize_advantage
+        self.normalize_advantage_exclude_dims = normalize_advantage_exclude_dims
+
         if gamma is not None:
             raise TypeError(_GAMMA_LMBDA_DEPREC_ERROR)
         self._set_deprecated_ctor_keys(
@@ -606,9 +614,16 @@ def forward(self, tensordict: TensorDictBase) -> TensorDictBase:
             )
             advantage = tensordict.get(self.tensor_keys.advantage)
         if self.normalize_advantage and advantage.numel() > 1:
-            loc = advantage.mean()
-            scale = advantage.std().clamp_min(1e-6)
-            advantage = (advantage - loc) / scale
+            if advantage.numel() > tensordict.batch_size.numel() and not len(
+                self.normalize_advantage_exclude_dims
+            ):
+                warnings.warn(
+                    "You requested advantage normalization and the advantage key has more dimensions"
+                    " than the tensordict batch. Make sure to pass `normalize_advantage_exclude_dims` "
+                    "if you want to keep any dimension independent while computing normalization statistics. "
+                    "If you are working in multi-agent/multi-objective settings this is highly suggested."
+                )
+            advantage = _standardize(advantage, self.normalize_advantage_exclude_dims)
 
         log_weight, dist, kl_approx = self._log_weight(tensordict)
         if is_tensor_collection(log_weight):
@@ -711,6 +726,9 @@ class ClipPPOLoss(PPOLoss):
             Can be one of "l1", "l2" or "smooth_l1". Defaults to ``"smooth_l1"``.
         normalize_advantage (bool, optional): if ``True``, the advantage will be normalized
             before being used. Defaults to ``False``.
+        normalize_advantage_exclude_dims (Tuple[int], optional): dimensions to exclude from the advantage
+            standardization. Negative dimensions are valid. This is useful in multiagent (or multiobjective) settings
+            where the agent (or objective) dimension may be excluded from the reductions. Default: ().
         separate_losses (bool, optional): if ``True``, shared parameters between
             policy and critic will only be trained on the policy loss.
             Defaults to ``False``, i.e., gradients are propagated to shared
@@ -802,6 +820,7 @@ def __init__(
         critic_coef: float = 1.0,
         loss_critic_type: str = "smooth_l1",
         normalize_advantage: bool = False,
+        normalize_advantage_exclude_dims: Tuple[int] = (),
         gamma: float = None,
         separate_losses: bool = False,
         reduction: str = None,
@@ -821,6 +840,7 @@ def __init__(
             critic_coef=critic_coef,
             loss_critic_type=loss_critic_type,
             normalize_advantage=normalize_advantage,
+            normalize_advantage_exclude_dims=normalize_advantage_exclude_dims,
             gamma=gamma,
             separate_losses=separate_losses,
             reduction=reduction,
@@ -871,9 +891,16 @@ def forward(self, tensordict: TensorDictBase) -> TensorDictBase:
             )
             advantage = tensordict.get(self.tensor_keys.advantage)
         if self.normalize_advantage and advantage.numel() > 1:
-            loc = advantage.mean()
-            scale = advantage.std().clamp_min(1e-6)
-            advantage = (advantage - loc) / scale
+            if advantage.numel() > tensordict.batch_size.numel() and not len(
+                self.normalize_advantage_exclude_dims
+            ):
+                warnings.warn(
+                    "You requested advantage normalization and the advantage key has more dimensions"
+                    " than the tensordict batch. Make sure to pass `normalize_advantage_exclude_dims` "
+                    "if you want to keep any dimension independent while computing normalization statistics. "
+                    "If you are working in multi-agent/multi-objective settings this is highly suggested."
+                )
+            advantage = _standardize(advantage, self.normalize_advantage_exclude_dims)
 
         log_weight, dist, kl_approx = self._log_weight(tensordict)
         # ESS for logging
@@ -955,6 +982,9 @@ class KLPENPPOLoss(PPOLoss):
             Can be one of "l1", "l2" or "smooth_l1". Defaults to ``"smooth_l1"``.
         normalize_advantage (bool, optional): if ``True``, the advantage will be normalized
             before being used. Defaults to ``False``.
+        normalize_advantage_exclude_dims (Tuple[int], optional): dimensions to exclude from the advantage
+            standardization. Negative dimensions are valid. This is useful in multiagent (or multiobjective) settings
+            where the agent (or objective) dimension may be excluded from the reductions. Default: ().
         separate_losses (bool, optional): if ``True``, shared parameters between
             policy and critic will only be trained on the policy loss.
             Defaults to ``False``, i.e., gradients are propagated to shared
@@ -1048,6 +1078,7 @@ def __init__(
         critic_coef: float = 1.0,
         loss_critic_type: str = "smooth_l1",
         normalize_advantage: bool = False,
+        normalize_advantage_exclude_dims: Tuple[int] = (),
         gamma: float = None,
         separate_losses: bool = False,
         reduction: str = None,
@@ -1063,6 +1094,7 @@ def __init__(
             critic_coef=critic_coef,
             loss_critic_type=loss_critic_type,
             normalize_advantage=normalize_advantage,
+            normalize_advantage_exclude_dims=normalize_advantage_exclude_dims,
             gamma=gamma,
             separate_losses=separate_losses,
             reduction=reduction,
@@ -1151,9 +1183,17 @@ def forward(self, tensordict: TensorDictBase) -> TensorDict:
             )
             advantage = tensordict_copy.get(self.tensor_keys.advantage)
         if self.normalize_advantage and advantage.numel() > 1:
-            loc = advantage.mean()
-            scale = advantage.std().clamp_min(1e-6)
-            advantage = (advantage - loc) / scale
+            if advantage.numel() > tensordict.batch_size.numel() and not len(
+                self.normalize_advantage_exclude_dims
+            ):
+                warnings.warn(
+                    "You requested advantage normalization and the advantage key has more dimensions"
+                    " than the tensordict batch. Make sure to pass `normalize_advantage_exclude_dims` "
+                    "if you want to keep any dimension independent while computing normalization statistics. "
+                    "If you are working in multi-agent/multi-objective settings this is highly suggested."
+                )
+            advantage = _standardize(advantage, self.normalize_advantage_exclude_dims)
+
         log_weight, dist, kl_approx = self._log_weight(tensordict_copy)
         neg_loss = log_weight.exp() * advantage