__all__ = ["SupervisedLearningProblem"]
import warnings
from typing import Dict, Tuple
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from evox.core import Problem, use_state, vmap
from .utils import get_vmap_model_state_forward
[文档]
class SupervisedLearningProblem(Problem):
"""The supervised learning problem to test a model's parameters or a batch of parameters with given data and criterion."""
def __init__(
self,
model: nn.Module,
data_loader: DataLoader,
criterion: nn.Module,
n_batch_per_eval: int = 1,
pop_size: int | None = None,
device: torch.device | None = None,
reduction: str = "mean",
):
"""Initialize the `SupervisedLearningProblem`.
:param model: The neural network model whose parameters need to be evaluated.
:param data_loader: The data loader providing the dataset for evaluation.
:param criterion: The loss function used to evaluate the parameters' performance.
:param n_batch_per_eval: The number of batches to be evaluated in a single evaluation. When set to -1, will go through the whole dataset. Defaults to 1.
:param pop_size: The size of the population (batch size of the parameters) to be evaluated. Defaults to None for single-run mode.
:param device: The device to run the computations on. Defaults to the current default device.
:param reduction: The reduction method for the criterion. 'mean' | 'sum'. Defaults to "mean".
:raises RuntimeError: If the data loader contains no items.
```{warning}
This problem does NOT support HPO wrapper (`problems.hpo_wrapper.HPOProblemWrapper`), i.e., the workflow containing this problem CANNOT be vmapped.
```
"""
super().__init__()
self.device = torch.get_default_device() if device is None else device
pop_size = 1 if pop_size is None else pop_size
self.data_loader = data_loader
self.data_loader_iter = iter(data_loader)
self.n_batch_per_eval = n_batch_per_eval
self.reduction = reduction
# JITed model state forward initialization
self.vmap_init_state, self.vmap_state_forward = get_vmap_model_state_forward(
model,
pop_size,
device=device,
)
self.init_state = model.state_dict()
self.state_forward = use_state(model)
# JITed and vmapped state criterion initialization
self.criterion = criterion
self.vmap_criterion = vmap(criterion, in_dims=(0, None))
[文档]
def _vmap_forward_pass(
self, model_state: Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor]], data: Tuple[torch.Tensor, torch.Tensor]
):
inputs, labels = data
inputs = inputs.to(device=self.device, non_blocking=True)
labels = labels.to(device=self.device, non_blocking=True)
_model_state, logits = self.vmap_state_forward(
model_state,
inputs,
)
loss = self.vmap_criterion(
logits,
labels,
)
return loss
[文档]
def _forward_pass(
self, model_state: Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor]], data: Tuple[torch.Tensor, torch.Tensor]
):
inputs, labels = data
inputs = inputs.to(device=self.device, non_blocking=True)
labels = labels.to(device=self.device, non_blocking=True)
_model_state, logits = self.state_forward(
model_state,
inputs,
)
loss = self.criterion(
logits,
labels,
)
return loss
[文档]
def _vmap_evaluate(
self,
pop_params: Dict[str, nn.Parameter],
):
# Initialize model and criterion states
model_state = self.vmap_init_state | pop_params
losses = []
if self.n_batch_per_eval == -1:
for data in self.data_loader:
losses.append(self._vmap_forward_pass(model_state, data))
else:
for _ in range(self.n_batch_per_eval):
losses.append(self._vmap_forward_pass(model_state, next(self.data_loader_iter)))
if losses[0].ndim == 1:
losses = torch.stack(losses, dim=1)
warnings.warn(
"Criterion output is a scalar. We recommend setting `reduction` to 'none' for criterion and let SupervisedLearningProblem handle the reduction."
)
else:
losses = torch.cat(losses, dim=1)
if self.reduction == "mean":
pop_fitness = losses.mean(dim=1)
elif self.reduction == "sum":
pop_fitness = losses.sum(dim=1)
return pop_fitness
[文档]
def _single_evaluate(
self,
params: Dict[str, nn.Parameter],
):
# Initialize model and criterion states
model_state = self.init_state | params
losses = []
if self.n_batch_per_eval == -1:
for data in self.data_loader:
losses.append(self._forward_pass(model_state, data))
else:
for _ in range(self.n_batch_per_eval):
losses.append(self._forward_pass(model_state, next(self.data_loader_iter)))
if losses[0].ndim == 0:
losses = torch.stack(losses, dim=0)
warnings.warn(
"Criterion output is a scalar. We recommend setting `reduction` to 'none' for criterion and let SupervisedLearningProblem handle the reduction."
)
else:
losses = torch.cat(losses, dim=0)
if self.reduction == "mean":
pop_fitness = losses.mean(dim=0)
elif self.reduction == "sum":
pop_fitness = losses.sum(dim=0)
return pop_fitness
[文档]
def evaluate(self, pop_params: Dict[str, nn.Parameter]) -> torch.Tensor:
"""Evaluate the fitness of a population (batch) of model parameters.
:param pop_params: A dictionary of parameters where each key is a parameter name and each value is a tensor of shape (batch_size, *param_shape) representing the batched parameters of batched models.
:return: A tensor of shape (batch_size,) containing the fitness of each sample in the population.
"""
pop_size = next(iter(pop_params.values())).size(0)
if pop_size != 1:
pop_fitness = self._vmap_evaluate(pop_params)
else:
pop_fitness = self._single_evaluate(pop_params)
return pop_fitness
EvoX