使用 state_dict 保存与加载
保存权重检查点，以便稍后继续训练
使用 state_dict 保存与加载 是 CoddyKit 上的免费 Deep Learning Academy 课时。 这是第 3 节课，共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 Deep Learning Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Deep Learning Academy 课程共包含 4 节课。
本课时的部分内容尚未翻译，以英文显示。
Why Checkpoints Matter
Training can take hours, and crashes happen. Saving your progress as a checkpoint lets you stop, resume, or ship the model without retraining from zero.
What a state_dict Holds
A model's state_dict is a plain dictionary mapping each layer name to its learned tensors. It is everything the model knows, packed for storage.
model.state_dict()Save the Weights
Use torch.save on the state_dict to write the weights to disk. The .pt or .pth extension is the common convention for these files.
torch.save(model.state_dict(), 'model.pt')Load the Weights Back
To restore, read the file with torch.load and pour it into a model using load_state_dict. The architecture must match the saved one.
model.load_state_dict(torch.load('model.pt'))Recreate the Architecture First
A state_dict holds numbers, not the class itself. You must build the same model object in code before you can load weights into it.
model = MyNet()
model.load_state_dict(torch.load('model.pt'))Eval Mode After Loading
Right after loading for inference, call model.eval(). It switches dropout and batch norm into prediction behavior so outputs are correct.
model.eval()Save the Optimizer Too
To truly resume training, also save the optimizer's state_dict. It holds momentum and adaptive stats that would otherwise reset to zero.
torch.save(optimizer.state_dict(), 'opt.pt')Bundle a Full Checkpoint
Pack model, optimizer, and the current epoch into one checkpoint dict. Now a single file restores your entire training session.
ckpt = {'epoch': epoch, 'model': model.state_dict(), 'opt': optimizer.state_dict()}
torch.save(ckpt, 'ckpt.pt')Resume From a Checkpoint
Load the bundle and restore each piece in turn. Reading the saved epoch lets you continue the loop exactly where it left off.
ckpt = torch.load('ckpt.pt')
model.load_state_dict(ckpt['model'])
optimizer.load_state_dict(ckpt['opt'])Map to the Right Device
If you saved on GPU and load on CPU, pass map_location to torch.load. It moves the weights to a device your machine actually has.
torch.load('model.pt', map_location='cpu')Save the Best, Not the Last
Watch validation loss and overwrite your checkpoint only when it improves. That way you keep the best model, not whatever the final epoch produced.
Quick Check
You saved only model.state_dict(). What must exist before you can load it?
Recap
Save weights with state_dict and torch.save, rebuild the model to load them, and bundle the optimizer and epoch for a full resume. Keep the best one. 💾
用 AI 导师学习 Python — 免费
在浏览器中编写并运行真实代码，获得全天候 AI 导师的即时帮助，并在网页或应用中继续学习。
- 课程
- 30
- 课程
- 120
常见问题解答
「使用 state_dict 保存与加载」课时是免费的吗？
是的 — 「使用 state_dict 保存与加载」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Deep Learning Academy 课程的其余内容，请升级到 CoddyKit PRO。 Deep Learning Academy 课程共包含 4 节课。
「使用 state_dict 保存与加载」这节课中我会学到什么？
保存权重检查点，以便稍后继续训练 你通过在浏览器中直接运行的动手代码来练习 Deep Learning Academy，全天候 AI 导师会在你学习这节课的过程中回答你的问题。
学习 Deep Learning Academy 需要有经验吗？
无需任何先前经验。CoddyKit 上的 Deep Learning Academy 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。 这是第 3 节课，共 4 节。
「使用 state_dict 保存与加载」课时需要多长时间？
大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。
我能在这节 Deep Learning Academy 课中编写并运行代码吗？
能。每节 Deep Learning Academy 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。
此课程中的所有课时
- 划分训练集、验证集与测试集
- 带验证的周期循环
- 使用 state_dict 保存与加载
- 根据验证损失提前停止