import torch.nn as nn
from pydantic import Field, model_validator
from typing_extensions import Self
from typing import Literal, List, Union, Dict
from onyxengine.modeling import (
OnyxModelBaseConfig,
OnyxModelOptBaseConfig,
validate_param,
validate_opt_param,
ModelSimulator,
FeatureScaler
)
[docs]
class MLPConfig(OnyxModelBaseConfig):
"""
Configuration class for the MLP model.
Args:
type (str): Model type = 'mlp', immutable.
outputs (List[Output]): List of output variables.
inputs (List[Input | State]): List of input variables.
dt (float): Time step for the model.
sequence_length (int): Length of the input sequence (default is 1).
hidden_layers (int): Number of hidden layers (default is 2).
hidden_size (int): Size of each hidden layer (default is 32).
activation (Literal['relu', 'gelu', 'tanh', 'sigmoid']): Activation function (default is 'relu').
dropout (float): Dropout rate for layers (default is 0.0).
bias (bool): Whether to use bias in layers (default is True).
"""
type: Literal['mlp'] = Field(default='mlp', frozen=True, init=False)
hidden_layers: int = 2
hidden_size: int = 32
activation: Literal['relu', 'gelu', 'tanh', 'sigmoid'] = 'relu'
dropout: float = 0.0
bias: bool = True
@model_validator(mode='after')
def validate_hyperparameters(self) -> Self:
validate_param(self.hidden_layers, 'hidden_layers', min_val=1, max_val=10)
validate_param(self.hidden_size, 'hidden_size', min_val=1, max_val=1024)
validate_param(self.dropout, 'dropout', min_val=0.0, max_val=1.0)
return self
class MLPOptConfig(OnyxModelOptBaseConfig):
"""
Optimization config class for the MLP model.
Args:
type (str): Model type = 'mlp_opt', immutable.
outputs (List[Output]): List of output variables.
inputs (List[Input | State]): List of input variables.
dt (float): Time step for the model.
sequence_length (Union[int, Dict[str, List[int]]): Length of the input sequence (default is {"select": [1, 2, 4, 5, 6, 8, 10]}).
hidden_layers (Union[int, Dict[str, List[int]]): Number of hidden layers (default is {"range": [2, 5, 1]}).
hidden_size (Union[int, Dict[str, List[int]]): Size of each hidden layer (default is {"select": [12, 24, 32, 64, 128]}).
activation (Union[Literal['relu', 'gelu', 'tanh', 'sigmoid'], Dict[str, List[str]]): Activation function (default is {"select": ['relu', 'gelu', 'tanh']}).
dropout (Union[float, Dict[str, List[float]]): Dropout rate for layers (default is {"range": [0.0, 0.4, 0.1]}).
bias (Union[bool, Dict[str, List[bool]]): Whether to use bias in layers (default is True).
"""
type: Literal['mlp_opt'] = Field(default='mlp_opt', frozen=True, init=False)
hidden_layers: Union[int, Dict[str, List[int]]] = {"range": [2, 5, 1]}
hidden_size: Union[int, Dict[str, List[int]]] = {"select": [12, 24, 32, 64, 128]}
activation: Union[Literal['relu', 'gelu', 'tanh', 'sigmoid'], Dict[str, List[str]]] = {"select": ['relu', 'gelu', 'tanh']}
dropout: Union[float, Dict[str, List[float]]] = {"range": [0.0, 0.4, 0.1]}
bias: Union[bool, Dict[str, List[bool]]] = True
@model_validator(mode='after')
def validate_hyperparameters(self) -> Self:
validate_opt_param(self.hidden_layers, 'hidden_layers', options=['select', 'range'], min_val=1, max_val=10)
validate_opt_param(self.hidden_size, 'hidden_size', options=['select', 'range'], min_val=1, max_val=1024)
validate_opt_param(self.activation, 'activation', options=['select'], select_from=['relu', 'gelu', 'tanh', 'sigmoid'])
validate_opt_param(self.dropout, 'dropout', options=['select', 'range'], min_val=0.0, max_val=1.0)
validate_opt_param(self.bias, 'bias', options=['select'], select_from=[True, False])
return self
class MLP(nn.Module, ModelSimulator):
def __init__(self, config: MLPConfig):
nn.Module.__init__(self)
ModelSimulator.__init__(
self,
outputs=config.outputs,
inputs=config.inputs,
sequence_length=config.sequence_length,
dt=config.dt,
)
self.feature_scaler = FeatureScaler(outputs=config.outputs, inputs=config.inputs)
self.config = config
num_inputs = len(config.inputs) * config.sequence_length
num_outputs = len(config.outputs)
hidden_layers = config.hidden_layers
hidden_size = config.hidden_size
activation = None
if config.activation == 'relu':
activation = nn.ReLU()
elif config.activation == 'gelu':
activation = nn.GELU()
elif config.activation == 'tanh':
activation = nn.Tanh()
elif config.activation == 'sigmoid':
activation = nn.Sigmoid()
else:
raise ValueError(f"Activation function {config.activation} not supported")
dropout = config.dropout
bias = config.bias
layers = []
# Add first hidden layer
layers.append(nn.Linear(num_inputs, hidden_size, bias=bias))
layers.append(nn.LayerNorm(hidden_size))
layers.append(activation)
layers.append(nn.Dropout(dropout))
# Add remaining hidden layers
for _ in range(hidden_layers-1):
layers.append(nn.Linear(hidden_size, hidden_size, bias=bias))
layers.append(nn.LayerNorm(hidden_size))
layers.append(activation)
layers.append(nn.Dropout(dropout))
# Add output layer
layers.append(nn.Linear(hidden_size, num_outputs, bias=bias))
self.model = nn.Sequential(*layers)
def forward(self, x):
# Sequence input shape (batch_size, sequence_length, num_inputs)
# Flatten to (batch_size, sequence_length * num_inputs)
x = self.feature_scaler.scale_inputs(x)
x = x.view(x.size(0), -1)
return self.feature_scaler.descale_outputs(self.model(x))