# SPDX-FileCopyrightText: Copyright (c) 2025 The ProtoMotions Developers
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
"""Common neural network components and utilities for agents.
This module provides shared building blocks used across different agent architectures,
including observation normalization, weight initialization, and specialized layers.
Key Classes:
- NormObsBase: Base class for modules with observation normalization
- ObsProcessor: General observation processor for reshaping and normalizing
- Embedding: Embedding layer for discrete inputs
Key Functions:
- weight_init: Initialize network weights
- get_params: Extract parameters from optimizer groups
"""
import torch
from torch import nn, Tensor
from protomotions.utils.hydra_replacement import get_class
from copy import copy
from typing import List, Dict, Optional
from tensordict import TensorDict
from tensordict.nn import TensorDictModuleBase
from protomotions.agents.utils.normalization import RunningMeanStd
from protomotions.agents.common.config import (
NormObsBaseConfig,
ObsProcessorConfig,
ModuleContainerConfig,
ModuleOperationConfig,
ModuleOperationForwardConfig,
ModuleOperationPermuteConfig,
ModuleOperationReshapeConfig,
ModuleOperationSqueezeConfig,
ModuleOperationUnsqueezeConfig,
ModuleOperationExpandConfig,
ModuleOperationSphereProjectionConfig,
)
[docs]
def get_params(obj) -> List[nn.Parameter]:
"""Extract parameters from optimizer parameter groups.
Handles both flat lists of parameters and grouped parameters (as used by optimizers).
Args:
obj: Either a list of nn.Parameter or a list of parameter groups (dicts).
Returns:
Flat list of all parameters.
"""
as_list = list(obj)
if isinstance(as_list[0], Tensor):
return as_list
else:
params = []
for group in as_list:
params = params + list(group["params"])
return params
[docs]
def weight_init(m, orthogonal=False):
"""Initialize weights for neural network modules.
Applies appropriate initialization to linear layers and other modules.
Linear layers get orthogonal or default initialization with zero bias.
Args:
m: Neural network module to initialize.
orthogonal: If True, use orthogonal initialization for linear layers.
"""
if isinstance(m, nn.Linear):
if orthogonal:
nn.init.orthogonal_(m.weight.data)
if hasattr(m.bias, "data"):
m.bias.data.fill_(0.0)
elif hasattr(m, "reset_parameters"):
m.reset_parameters()
[docs]
class NormObsBase(nn.Module):
"""Base class for modules with observation normalization.
Provides running mean/std normalization of observations using exponential
moving averages. Normalization statistics are updated during training and
frozen during evaluation.
Uses lazy initialization - input shape is inferred on first forward pass.
This is a simple tensor-to-tensor module. Subclasses handle TensorDict extraction/insertion.
Args:
config: Configuration specifying output dimensions and normalization parameters.
Attributes:
running_obs_norm: RunningMeanStd module for observation normalization (lazy).
num_out: Output dimension.
"""
[docs]
def __init__(self, config: NormObsBaseConfig):
super().__init__()
self.config = config
self.build_norm()
[docs]
def build_norm(self):
if self.config.normalize_obs:
# Use lazy initialization - shape inferred on first forward
self.running_obs_norm = RunningMeanStd(
fabric=None,
shape=None, # Lazy: inferred from first input
device="cpu",
clamp_value=self.config.norm_clamp_value,
)
[docs]
def forward(self, obs: Tensor) -> Tensor:
"""Forward pass that normalizes observations.
Args:
obs: Observation tensor to normalize.
Returns:
Normalized observation tensor.
"""
if self.config.normalize_obs:
norm_obs = self.running_obs_norm.normalize(obs)
if self.training:
self.running_obs_norm.record_moments(obs)
else:
norm_obs = obs
return norm_obs
[docs]
def apply_module_operations(
obs: Tensor,
module_operations: List[ModuleOperationConfig],
normalizer: NormObsBase,
forward_model: Optional[nn.Module] = None,
) -> Dict[str, Tensor]:
batch_size = obs.shape[0]
norm_obs = None
for operation in module_operations:
if isinstance(operation, ModuleOperationPermuteConfig):
obs = obs.permute(*operation.new_order)
elif isinstance(operation, ModuleOperationReshapeConfig):
new_shape = copy(operation.new_shape)
if isinstance(new_shape[0], str) and new_shape[0] == "batch_size":
# Use actual tensor dimension for ONNX tracing compatibility
new_shape[0] = batch_size
elif isinstance(new_shape[0], str) and "batch_size" in new_shape[0]:
# Handle expressions like "batch_size * 2" - only works for concrete values
try:
new_shape[0] = eval(new_shape[0].replace("batch_size", str(int(batch_size))))
except (TypeError, ValueError):
# During tracing, batch_size may not be convertible to int
# Fall back to using -1 which PyTorch will infer
new_shape[0] = -1
obs = obs.reshape(*new_shape)
elif isinstance(operation, ModuleOperationSqueezeConfig):
obs = obs.squeeze(dim=operation.squeeze_dim)
elif isinstance(operation, ModuleOperationUnsqueezeConfig):
obs = obs.unsqueeze(dim=operation.unsqueeze_dim)
elif isinstance(operation, ModuleOperationExpandConfig):
obs = obs.expand(*operation.expand_shape)
elif isinstance(operation, ModuleOperationSphereProjectionConfig):
obs = torch.nn.functional.normalize(obs, dim=-1)
elif isinstance(operation, ModuleOperationForwardConfig):
obs_shape = obs.shape
if len(obs_shape) > 2:
obs = obs.reshape(-1, obs_shape[-1])
if normalizer is not None:
obs = norm_obs = normalizer(obs)
if len(obs_shape) > 2:
obs = obs.reshape(*obs_shape[:-1], -1)
if norm_obs is not None:
norm_obs = norm_obs.reshape(*obs_shape[:-1], -1)
if forward_model is not None:
obs = forward_model(obs)
else:
raise NotImplementedError(f"Operation {operation} not implemented")
return_dict = {"output": obs}
if norm_obs is not None:
return_dict["norm_obs"] = norm_obs
return return_dict
[docs]
class ObsProcessor(TensorDictModuleBase):
"""General observation processor - applies operations and normalization.
ForwardConfig applies normalization but skips the forward model (no MLP).
"""
config: ObsProcessorConfig
[docs]
def __init__(self, config: ObsProcessorConfig):
TensorDictModuleBase.__init__(self)
self.config = config
assert len(config.out_keys) == 1, "ObsProcessor requires exactly one output key"
self.norm = NormObsBase(config) if config.normalize_obs else None
self.in_keys = config.in_keys
self.out_keys = config.out_keys
[docs]
def forward(self, tensordict: TensorDict, *args, **kwargs) -> TensorDict:
obs = torch.cat([tensordict[key] for key in self.in_keys], dim=-1)
result = apply_module_operations(
obs, self.config.module_operations, forward_model=None, normalizer=self.norm
)
tensordict[self.out_keys[0]] = result["output"]
return tensordict
[docs]
class ModuleContainer(TensorDictModuleBase):
"""Generic container that runs a list of modules sequentially on a TensorDict.
With TensorDict, the distinction between "sequential", "parallel input",
and "parallel output" is implicit in how keys flow between modules.
"""
config: ModuleContainerConfig
[docs]
def __init__(self, config: ModuleContainerConfig):
TensorDictModuleBase.__init__(self)
self.config = config
self.models = nn.ModuleList()
# Build models and validate data flow
# Container in_keys are "available" at the start (like sources)
available_keys = set(self.config.in_keys)
for i, model_cfg in enumerate(config.models):
model = get_class(model_cfg._target_)(config=model_cfg)
self.models.append(model)
# Validate: each model's in_keys must be available
# (either from container in_keys or a previous model's out_keys)
for in_key in model.in_keys:
assert in_key in available_keys, (
f"ModuleContainer: model {i} ({type(model).__name__}) requires "
f"input key '{in_key}' but it's not available.\n"
f" Available keys: {sorted(available_keys)}\n"
f" Model in_keys: {model.in_keys}"
)
# Add this model's outputs to available keys for subsequent models
available_keys.update(model.out_keys)
# Validate: every container out_key must have been produced
for out_key in self.config.out_keys:
assert out_key in available_keys, (
f"ModuleContainer: container promises out_key '{out_key}' "
f"but no model produces it.\n"
f" Available keys after all models: {sorted(available_keys)}"
)
self.in_keys = self.config.in_keys
self.out_keys = self.config.out_keys
[docs]
def forward(self, tensordict: TensorDict, *args, **kwargs) -> TensorDict:
"""Forward through all models sequentially."""
for model in self.models:
tensordict = model(tensordict)
return tensordict
