我该如何做早期停止;在Tsai的Fit（）期间
以下是我的示例代码来自
from tsai.all import *
dsid = 'AppliancesEnergy'
arch_config = {
'hidden_size':100,
'n_layers':2,
'rnn_dropout':0.2,
'fc_dropout':0.5,
'bidirectional':True
}
X, y, splits = get_regression_data(dsid, split_data=False)
learn = TSRegressor(
X,
y,
splits=splits,
bs=128,
batch_tfms=[TSStandardize(by_sample=True)],
arch=LSTM,
arch_config=arch_config,
metrics=[mae, rmse],
cbs=ShowGraph(),
verbose=True)
learn.fit_one_cycle(100, lr_max=1e-3)
learn.plot_metrics()
这很好。我想做的是在fit期间早点停止（）。 我在Fastai中找到了回调函数'terminateOnnancallback（）'，我在下面的Import fastai上应用了它。
learn.fit_one_cycle(100, lr_max=1e-3, cbs=TerminateOnNaNCallback())
但这不起作用。如果有人知道，请告诉我。 谢谢。
Below is my sample code from tsai notebook for time series regression problem.
from tsai.all import *
dsid = 'AppliancesEnergy'
arch_config = {
'hidden_size':100,
'n_layers':2,
'rnn_dropout':0.2,
'fc_dropout':0.5,
'bidirectional':True
}
X, y, splits = get_regression_data(dsid, split_data=False)
learn = TSRegressor(
X,
y,
splits=splits,
bs=128,
batch_tfms=[TSStandardize(by_sample=True)],
arch=LSTM,
arch_config=arch_config,
metrics=[mae, rmse],
cbs=ShowGraph(),
verbose=True)
learn.fit_one_cycle(100, lr_max=1e-3)
learn.plot_metrics()
This works good. What I want to do is Early Stopping during fit().
I found callbacks function 'TerminateOnNaNCallback()' for that in fastai and I applied it like below with import fastai.
learn.fit_one_cycle(100, lr_max=1e-3, cbs=TerminateOnNaNCallback())
But this does not work. If somebody knows, please let me know.
Thank you.
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（1）
您只需导入并使用早期停止回调来自fastai with fastai with tsai库：
然后设置回调：
您可以定义参数，例如监视哪个度量/损失，以及在未经改进的情况下终止了多少转。
You can just import and use the early stopping callback from fastai with the tsai library:
Then set your callbacks:
You can define parameters such as which metric/loss is monitored and after how many turns of no improvement the training is terminated.