Agent Skills
› aiming-lab/AutoResearchClaw
› pytorch-training
pytorch-training
GitHub提供构建健壮 PyTorch 训练循环的最佳实践，涵盖可复现性、数据加载优化、学习率调度及模型检查点等。用于生成或审查 ML 训练代码。
触发场景
生成 PyTorch 训练代码
审查 ML 训练代码
安装
npx skills add aiming-lab/AutoResearchClaw --skill pytorch-training -g -y
SKILL.md
Frontmatter
{
"name": "pytorch-training",
"metadata": {
"author": "researchclaw",
"version": "1.0",
"category": "tooling",
"priority": "3",
"references": "PyTorch Performance Tuning Guide, pytorch.org",
"code-template": "import torch\nimport torch.nn as nn\nfrom torch.utils.data import DataLoader\n\n# Reproducibility\ntorch.manual_seed(seed)\ntorch.cuda.manual_seed_all(seed)\ntorch.backends.cudnn.deterministic = True\n\n# Training loop\nmodel.train()\nfor epoch in range(num_epochs):\n for batch in train_loader:\n optimizer.zero_grad(set_to_none=True)\n loss = criterion(model(batch['input']), batch['target'])\n loss.backward()\n torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)\n optimizer.step()\n scheduler.step()",
"trigger-keywords": "training,pytorch,torch,deep learning,neural network,model",
"applicable-stages": "10,12"
},
"description": "Best practices for building robust PyTorch training loops. Use when generating or reviewing ML training code."
}
PyTorch Training Best Practice
- Use torch.manual_seed() for reproducibility (set for torch, numpy, random)
- Use DataLoader with num_workers>0 and pin_memory=True for GPU
- Enable cudnn.benchmark=True for fixed input sizes
- Use learning rate schedulers (CosineAnnealingLR or OneCycleLR)
- Implement early stopping based on validation metric
- Log metrics every epoch, save best model checkpoint
- Use torch.no_grad() for evaluation
- Clear gradients with optimizer.zero_grad(set_to_none=True) for efficiency
版本历史
- e2e23c9 当前 2026-07-25 07:48
