【Hackathon 10th Spring No.9】NewtonNet 复现

[co63oc]

PaddlePaddle/Paddle#77429
NewtonNet 复现

torch版本
https://github.com/co63oc/NewtonNet/tree/fix1 fix1分支
修改主要为配置修改，输出日志信息，保存权重，转换权重为paddle
配置修改训练批次大小50
转换权重脚本 https://github.com/co63oc/NewtonNet/blob/fix1/scripts/convert_paddle.py
运行脚本

cd scripts
python  newtonnet_train.py  -c config.yml

运行输出日志
https://github.com/co63oc/NewtonNet/blob/fix1/scripts/torch_newtonnet.log

当前PR 为paddle版本，需要安装paddle_geometric，paddle_scatter
配置文件 interatomic_potentials/configs/newtonnet/newtonnet.yaml
权重 interatomic_potentials/configs/newtonnet/newtonnet.pdparams，只是用来加载和torch一致的初始化参数，不是训练的结果权重
训练时需要使用数据集计算 stats_calc 更新权重，增加pretrained_need_update_by_train_loader配置，加载数据集后更新权重
增加 clip_grad 配置项，前向计算时需要修改模型参数
模型增加精度配置 precision，用来修改模型精度，修改模型精度时需要修改数据集精度，删除已有数据集 interatomic_potentials/example_data/aspirin/ccsd_train/processed/data.pt 运行后会重新生成
运行脚本

cd interatomic_potentials
python train.py -c configs/newtonnet/newtonnet.yaml

运行日志 https://github.com/co63oc/NewtonNet/blob/fix1/scripts/paddle_newtonnet.log

测试加载相同初始化参数，使用相同数据集按顺序读取训练，训练两轮以上loss一致，因为使用较多scatter计算，训练轮次逐渐增加时会增加浮点误差

torch

batch15 tensor(40.9183, device='cuda:0', dtype=torch.float64, grad_fn=<AddBackward0>)
batch16 tensor(34.6856, device='cuda:0', dtype=torch.float64, grad_fn=<AddBackward0>)
batch17 tensor(44.9074, device='cuda:0', dtype=torch.float64, grad_fn=<AddBackward0>)
batch18 tensor(52.8147, device='cuda:0', dtype=torch.float64, grad_fn=<AddBackward0>)
batch19 tensor(38.0302, device='cuda:0', dtype=torch.float64, grad_fn=<AddBackward0>) 

paddle

[2026/04/08 10:07:30] ppmat INFO: Train: Epoch [2/5] | Step: [15/19] | lr: 0.001000 | reader_cost: 0.000135 | batch_cost: 4.812305 | loss(loss): 40.918344�[0m
[2026/04/08 10:07:37] ppmat INFO: Train: Epoch [2/5] | Step: [16/19] | lr: 0.001000 | reader_cost: 0.000164 | batch_cost: 7.384202 | loss(loss): 34.685635�[0m
[2026/04/08 10:07:45] ppmat INFO: Train: Epoch [2/5] | Step: [17/19] | lr: 0.001000 | reader_cost: 0.000152 | batch_cost: 7.996556 | loss(loss): 44.907362�[0m
[2026/04/08 10:07:50] ppmat INFO: Train: Epoch [2/5] | Step: [18/19] | lr: 0.001000 | reader_cost: 0.000135 | batch_cost: 5.235708 | loss(loss): 52.814710�[0m
[2026/04/08 10:07:58] ppmat INFO: Train: Epoch [2/5] | Step: [19/19] | lr: 0.001000 | reader_cost: 0.000135 | batch_cost: 7.812943 | loss(loss): 38.030187�[0m

增加 paddle_save_fix，修改 paddle.save 调用会有权重丢失问题，复制权重后保存

[paddle-bot]

Thanks for your contribution!