技术科学(AI for Technology)
diff --git a/docs/index.md b/docs/index.md index ce3960db2d..acc20c6d44 100644 --- a/docs/index.md +++ b/docs/index.md @@ -79,6 +79,8 @@ | 微分方程 | [梯度增强的物理知识融合 PDE 求解](https://github.com/PaddlePaddle/PaddleScience/blob/develop/examples/gpinn/poisson_1d.py) | 机理驱动 | gPINN | 无监督学习 | - | [Paper](https://doi.org/10.1016/j.cma.2022.114823) | | 积分方程 | [沃尔泰拉积分方程](./zh/examples/volterra_ide.md) | 机理驱动 | MLP | 无监督学习 | - | [Project](https://github.com/lululxvi/deepxde/blob/master/examples/pinn_forward/Volterra_IDE.py) | | 微分方程 | [分数阶微分方程](https://github.com/PaddlePaddle/PaddleScience/blob/develop/examples/fpde/fractional_poisson_2d.py) | 机理驱动 | MLP | 无监督学习 | - | - | +| 光孤子 | [Optical soliton](./zh/examples/nlsmb.md) | 机理驱动 | MLP | 无监督学习 | - | [Paper](https://doi.org/10.1007/s11071-023-08824-w)| +| 光纤怪波 | [Optical rogue wave](./zh/examples/nlsmb.md) | 机理驱动 | MLP | 无监督学习 | - | [Paper](https://doi.org/10.1007/s11071-023-08824-w)|技术科学(AI for Technology)
diff --git a/ppsci/solver/eval.py b/ppsci/solver/eval.py index 186fe7bf6e..8087d45ced 100644 --- a/ppsci/solver/eval.py +++ b/ppsci/solver/eval.py @@ -63,6 +63,9 @@ def _eval_by_dataset( ) -> Tuple[float, Dict[str, Dict[str, float]]]: """Evaluate with computing metric on total samples(default process). + NOTE: This is the default evaluation method as general for most cases, but may not + memory-efficiency for large dataset or large output. + Args: solver (solver.Solver): Main Solver. epoch_id (int): Epoch id. @@ -159,6 +162,7 @@ def _eval_by_dataset( metric_dict_group: Dict[str, Dict[str, float]] = misc.PrettyOrderedDict() for metric_name, metric_func in _validator.metric.items(): + # NOTE: compute metric with entire output and label metric_dict = metric_func(all_output, all_label) metric_dict_group[metric_name] = { k: float(v) for k, v in metric_dict.items() @@ -189,6 +193,10 @@ def _eval_by_batch( ) -> Tuple[float, Dict[str, Dict[str, float]]]: """Evaluate with computing metric by batch, which is memory-efficient. + NOTE: This is a evaluation function for large dataset or large output, as is more + memory-efficiency than evaluating by dataset, but less general because some metric + is not independent among samples, e.g. L2 relative error. + Args: solver (solver.Solver): Main Solver. epoch_id (int): Epoch id. @@ -270,7 +278,10 @@ def _eval_by_batch( # concatenate all metric and discard metric of padded sample(s) for metric_name, metric_dict in metric_dict_group.items(): for var_name, metric_value in metric_dict.items(): + # NOTE: concat all metric(scalars) into metric vector metric_value = paddle.concat(metric_value)[:num_samples] + # NOTE: compute metric via averaging metric vector, + # this might be not general for certain evaluation case metric_value = float(metric_value.mean()) metric_dict_group[metric_name][var_name] = metric_value metric_str = f"{_validator.name}/{metric_name}.{var_name}"