运行pytorch_forecasting导入报错，pytorch_lightning.metrics已迁移至torchmetrics

问题说明 #

运行以下导入代码时，from pytorch_forecasting.metrics import MAE, SMAPE, PoissonLoss, QuantileLoss 语句报错，原因是pytorch_forecasting内部引用了已废弃的from pytorch_lightning.metrics import Metric as LightningMetric，而pytorch_lightning.metrics已迁移至torchmetrics，且无法直接编辑pytorch_forecasting模块，安装旧版本包也引发更多问题。

导入代码：

from pytorch_forecasting import Baseline, TemporalFusionTransformer, TimeSeriesDataSet
from pytorch_forecasting.data import GroupNormalizer
from pytorch_forecasting.metrics import MAE, SMAPE, PoissonLoss, QuantileLoss
from pytorch_forecasting.models.temporal_fusion_transformer.tuning import optimize_hyperparameters

解决方案 #

• 方法1：临时映射模块（无需修改原代码）
  在你的代码最开头添加以下内容，将pytorch_lightning.metrics的引用定向到torchmetrics：

import sys
from types import ModuleType

# 创建模拟pytorch_lightning.metrics的模块
pl_metrics = ModuleType("pytorch_lightning.metrics")
sys.modules["pytorch_lightning.metrics"] = pl_metrics

# 从torchmetrics导入Metric并赋值给模拟模块
from torchmetrics import Metric
pl_metrics.Metric = Metric

# 后续正常执行原导入语句
from pytorch_forecasting import Baseline, TemporalFusionTransformer, TimeSeriesDataSet
from pytorch_forecasting.data import GroupNormalizer
from pytorch_forecasting.metrics import MAE, SMAPE, PoissonLoss, QuantileLoss
from pytorch_forecasting.models.temporal_fusion_transformer.tuning import optimize_hyperparameters

• 方法2：安装兼容版本组合
  安装已验证可解决此依赖问题的版本组合（可根据自身torch版本微调）：

pip install pytorch-forecasting==0.10.3 torchmetrics==0.11.4

• 方法3：动态patch模块引用
  使用unittest.mock临时替换模块路径：

from unittest.mock import patch

# 在导入pytorch_forecasting前patch依赖
with patch("pytorch_forecasting.metrics.base.pytorch_lightning.metrics", __import__("torchmetrics")):
    from pytorch_forecasting import Baseline, TemporalFusionTransformer, TimeSeriesDataSet
    from pytorch_forecasting.data import GroupNormalizer
    from pytorch_forecasting.metrics import MAE, SMAPE, PoissonLoss, QuantileLoss
    from pytorch_forecasting.models.temporal_fusion_transformer.tuning import optimize_hyperparameters

内容的提问来源于stack exchange，提问作者Jasper